50 of 51 criteria have a sourced finding; 46 were also cross-checked by a second agent. Every comparison that includes kodo links here for scope, limitations, atomic answers and sources.
Humans approve the rubric, templates and prototype; they do not review every finding. An agent source cross-check is not a hands-on product test. A missing sourced finding is not evidence that a capability is absent.
Terminal CLI on Windows, macOS and Linux (CI matrix covers all three) plus a loopback local web dashboard and an HTML log viewer for inspecting runs; no desktop app, mobile app or editor extension.
Scope
Surfaces: kodo CLI (interactive and non-interactive), kodo dashboard (local loopback web UI), kodo log viewer (HTML) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
CLI-first; the dashboard is read-mostly (feedback and stop endpoints are TODO stubs) and local-only.
9 atomic answers
Desktop os [partial] No desktop app; the CLI runs on Windows, macOS and Linux (CI matrix windows-latest, macos-latest, ubuntu-latest, Python 3.13; README gives uv install for Windows PowerShell). Sources: kodo:ci, kodo:readme.
Mobile apps [unsupported] Bounded search of kodo/ and README for ios, android, mobile found nothing; no mobile app. Sources: kodo:srch-mobile.
Browser [partial] `kodo dashboard` serves a local browser UI (run list, events, stats, stage view) and `kodo logs` opens an HTML viewer; both run on the host machine. Sources: kodo:dash, kodo:dashjs, kodo:viewer.
Mobile browser [unsupported] Dashboard and viewer HTML include a viewport meta tag but dashboard.css has no @media rules, and the dashboard server is loopback-only, so a phone browser cannot reach it. Sources: kodo:dash, kodo:dashjs.
Editor extensions [unsupported] Bounded search found no IDE or editor extension; only the CLI entry point in pyproject. Sources: kodo:srch-mobile, kodo:pyproject.
Task actions [partial] CLI starts runs (goal, test, improve, fix-from) and resumes them; the dashboard cannot start runs, and its feedback/stop POST handlers only acknowledge (TODO). A human can steer via lines appended to human_feedback.txt when --coach is on; approve is not a kodo concept (agents run bypassPermissions). Sources: kodo:params, kodo:main, kodo:dash, kodo:coach, kodo:claudesess.
Notifications [unsupported] Bounded search found no notification mechanism; progress prints to the terminal table. Sources: kodo:srch-notify.
Cross device continuity [partial] Run state persists under ~/.kodo/runs on the host and `kodo --resume` continues from the same machine; no other-device access. Sources: kodo:log, kodo:main.
Seven session backends launched or driven locally: Claude Code (claude-agent-sdk), Cursor CLI, OpenAI Codex CLI, Gemini CLI, Kimi (ACP), Kiro and OpenCode; orchestrator model via pydantic-ai providers or a CLI backend. Mixed-harness teams are native.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: claude, cursor, codex, gemini-cli, kimi, kiro, opencode
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
5 atomic answers
Harnesses [supported] BACKENDS registers claude, cursor, codex, gemini-cli, kimi, kiro, opencode; README lists six (OpenCode registered in code). Sources: kodo:backends, kodo:readme.
Model providers [supported] API orchestrator aliases through pydantic-ai-slim (anthropic, google, openai, groq, mistral, xai) and local Ollama; worker models are chosen per team agent in each CLI. Sources: kodo:models, kodo:pyproject, kodo:providers.
Custom adapters [partial] Custom teams assign any registered backend and model per role via team.json; adding a new harness needs a Python session class (library callers can subclass Session). Sources: kodo:teamcfg, kodo:backends.
Session integration [supported] Claude via claude-agent-sdk, Kimi via ACP JSON-RPC, Codex/Cursor/Gemini/Kiro/OpenCode by launching their CLIs as subprocess sessions with resume IDs; the CLI orchestrators reach tools through an internal MCP bridge. Sources: kodo:claudesess, kodo:kimisess, kodo:codexsess, kodo:mcpserver.
Mixed harness operation [supported] Default full team mixes Cursor (worker_fast, tester) and Claude (worker_smart, architect); README documents mixed teams. Sources: kodo:teamfull, kodo:readme.
Reuses installed agent subscriptions (Claude Max/Pro, Cursor, ChatGPT, Gemini free tier, Kiro) through each native CLI login, supports API keys for the orchestrator, and local Ollama for the orchestrator only. No vendor credits.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
Not separately cross-checked
7 atomic answers
Agent subscriptions [supported] Claude workers run under the Claude subscription (ANTHROPIC_API_KEY stripped), Cursor via subscription, Codex via ChatGPT Plus/Pro, Kiro via AWS Builder ID, Kimi via native login. Sources: kodo:providers, kodo:env.
Api accounts [supported] Orchestrator uses ANTHROPIC_API_KEY, GOOGLE_API_KEY/GEMINI_API_KEY or OPENAI_API_KEY; Codex accepts an OpenAI API key; Gemini CLI a Gemini API key. Sources: kodo:providers, kodo:models.
Vendor credits [unsupported] kodo has no hosted service or credits; bounded search of README, docs and kodo/ found no purchase flow. Sources: kodo:readme, kodo:srch-budget.
Local models [partial] Ollama models can be the orchestrator (ollama:<model> against localhost:11434); worker backends still need a coding-agent CLI. Sources: kodo:providers, kodo:models.
Account switching [unsupported] No multi-account switching feature in kodo/; each CLI keeps its own login. Sources: kodo:srch-auth.
Vendor account requirement [supported] No kodo account exists; at least one worker backend and (for the recommended API orchestrator) a provider key are needed. Sources: kodo:readme, kodo:providers.
Provider restrictions [supported] Docs note per-backend requirements (Claude Max/Pro, Cursor cursor-agent enablement); the README labels subscription-covered cost as Virtual and warns that native Kimi reports no usage. No provider terms are summarised. Sources: kodo:providers, kodo:readme.
Install with `uv tool install kodo-agent`, install at least one agent backend CLI, set an orchestrator API key, then run `kodo` interactively or with --goal. Works on any directory; Git is optional but recommended and needed for parallel worktree stages.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
No managed-cloud or BYO-VPS onboarding exists; Git-less behaviour is inferred from code rather than documented.
7 atomic answers
First task setup [supported] README Install and Usage: install uv, `uv tool install kodo-agent`, install a backend and key, run `kodo --goal ... --project ...`. Sources: kodo:readme.
No git [partial] Run-status exclusion and worktrees check for a .git ancestor and skip otherwise (run_status.py), so plain directories run; parallel worktree stages need Git; README warns to commit first. Sources: kodo:runstatus, kodo:gitops, kodo:readme.
Local git without remote [supported] All Git work is local (worktrees, --no-ff merges, auto-commit that never pushes); no remote needed. Sources: kodo:gitops.
Hosted git [partial] No GitHub/GitLab API integration (no PR or clone flow); works in an existing clone of any host. Sources: kodo:srch-pr.
Prerequisite guidance [supported] README Prerequisites table and docs/providers.md give install, auth and troubleshooting per backend; `kodo backends` shows availability and API key status. Sources: kodo:readme, kodo:providers, kodo:subcmd.
Managed cloud setup [unsupported] No managed cloud offering; bounded search found no hosted runtime. Sources: kodo:srch-mobile, kodo:readme.
Byo vps onboarding [unsupported] No VPS installer or remote-host onboarding; the CLI is installed on whichever host runs it. Sources: kodo:srch-mobile.
Programmable through the CLI (--json, non-interactive flags, --resume, runs/logs subcommands), a Python library API, and the local dashboard REST/SSE routes; no external MCP server, webhooks or stable-API commitment.
Scope
Surfaces: kodo CLI (interactive and non-interactive), kodo dashboard (local loopback web UI) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
REST routes are internal to the dashboard; steer and stop endpoints are stubs; no auth.
8 atomic answers
Cli [supported] kodo --goal/--goal-file/--test/--improve/--fix-from/--resume/--json and subcommands runs, logs, dashboard, backends, teams. Sources: kodo:params, kodo:subcmd.
Api [partial] Dashboard serves GET /api/runs, /api/run/{id}, /events, /stream (SSE), /stats and PUT /stages; it is a local viewer API, not documented as public. Sources: kodo:dash.
Sdk [partial] Python package exposes Agent, sessions and orchestrators (agent.py docstring examples); no versioned SDK documentation. Sources: kodo:agent, kodo:srch-api.
Mcp server [unsupported] The team MCP server is built per run for CLI orchestrators only; bounded search found no MCP server exposing kodo to other agents. Sources: kodo:mcpserver, kodo:srch-api.
Webhooks [unsupported] No webhook support found. Sources: kodo:srch-sched.
Operations [partial] Create (CLI), inspect (runs/logs/API), stop via Ctrl-C only; steer is limited to human_feedback.txt with --coach; review means reading logs. Sources: kodo:params, kodo:coach, kodo:dash.
Authentication [unsupported] Dashboard has no authentication; it binds only to 127.0.0.1 via LoopbackHTTPServer. Sources: kodo:srch-auth, kodo:dash.
Stability [partial] pyproject classifier Development Status 4 - Beta, version 0.5.1; no API stability policy documented. Sources: kodo:pyproject.
kodo itself is an MCP client only through its backends; it adds an internal team MCP bridge for CLI orchestrators and has no plugin system. Team roles accept custom system prompts; instruction files are read natively by each backend.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
No kodo-managed plugin or tool marketplace; MCP and skills come from the underlying CLIs.
7 atomic answers
Mcp client [partial] kodo wires its own team MCP server into Claude Code, Codex, Cursor, Gemini and Kimi orchestrators; user MCP servers are whatever those CLIs already have configured. Sources: kodo:mcpserver, kodo:agenttools.
Plugins [unsupported] Bounded search found no plugin API; teams and prompts are JSON/Python only. Sources: kodo:srch-api.
Custom tools [partial] Custom agents via team.json (system_prompt, description); no arbitrary tool registry. Sources: kodo:teamcfg.
Instructions import [partial] kodo does not import instruction files itself except OpenCode (AGENTS.md retained when requested); Claude Code, Cursor and Codex read CLAUDE.md/AGENTS.md natively as they run in the project. Sources: kodo:opencode, kodo:readme.
Authentication [partial] Integration auth is inherited from each backend CLI; kodo documents `kimi login` and per-backend setup. Sources: kodo:providers.
Permission scope [partial] Agents run bypassPermissions (Claude) or workspace-write (Codex); no per-integration scope in kodo. Sources: kodo:claudesess, kodo:codexsess.
Integration owner [supported] Connections are owned by the underlying CLIs; the internal team bridge is owned by kodo. Sources: kodo:mcpserver, kodo:providers.
CLI is a Python 3.13 package (setuptools, uv tool install); the web dashboard is a plain HTML/CSS/JS page served by a stdlib HTTP server. No native, Electron, Tauri or mobile client.
Scope
Surfaces: kodo CLI (interactive and non-interactive), kodo dashboard (local loopback web UI) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
4 atomic answers
Desktop implementation [supported] No desktop client; terminal CLI (Python, questionary prompts) and local browser dashboard. Sources: kodo:pyproject, kodo:dash.
Mobile implementation [unsupported] No mobile client in repo. Sources: kodo:srch-mobile.
Hybrids [unsupported] No Electron/Tauri/React Native; dashboard.html/css/js served by http.server. Sources: kodo:dash, kodo:dashjs.
Work enters as a goal (inline text, goal file, interactive interview with optional Claude refinement) or as test/improve/fix-from modes. No issue-tracker import, Git or chat triggers, and no external tracker relationship.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
Prompt/file intake only; no tracker integration.
9 atomic answers
Local creation [supported] Goals are created locally via --goal, --goal-file or interactive mode; intake breaks a goal into stages with acceptance criteria. Sources: kodo:readme, kodo:types.
Prompt intake [supported] Interactive Claude interview refines goals; --auto-refine and --skip-intake for non-interactive runs. Sources: kodo:readme.
External import [unsupported] Bounded search found no Linear/GitHub/Jira import. Sources: kodo:srch-issues, kodo:srch-pr.
Chat git triggers [unsupported] No chat or Git event triggers. Sources: kodo:srch-sched.
Authoritative tracker [unsupported] kodo has no tracker; goal.md and the run directory are the only records. Sources: kodo:srch-issues.
Stable ids [partial] Runs have timestamp IDs (e.g. 20260218_205503) and stages have 1-based indices; no durable task IDs across runs. Sources: kodo:log, kodo:types.
Supported edits [partial] Stage name, description and acceptance criteria can be edited in goal-plan.json through the dashboard PUT route. Sources: kodo:dash.
External roundtrips [unsupported] No external tracker, so no round trips. Sources: kodo:srch-issues.
Intake produces a GoalPlan of ordered stages each with a description and acceptance criteria, stored as goal-plan.json per run; stages can be edited through the dashboard and advisors can append stages. No revision history.
Scope
Surfaces: kodo CLI (interactive and non-interactive), kodo dashboard (local loopback web UI) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
No revision history or human editor beyond the dashboard JSON merge; plans are per-run.
7 atomic answers
Plans [supported] GoalPlan with shared context and stages; resumed runs reload goal-plan.json. Sources: kodo:types, kodo:dash.
Editors [partial] Dashboard PUT /api/run/{id}/stages edits name, description and acceptance_criteria; interactive intake lets a human review the plan. Sources: kodo:dash.
Revisions [unsupported] The PUT handler overwrites goal-plan.json in place; no versioning found. Sources: kodo:dash.
Implementation links [partial] Auto-commit messages and run logs tie stage results to commits, but plan stages carry no commit links. Sources: kodo:gitops, kodo:log.
Amendment [partial] Dashboard edits and the advisor can change stages between cycles; mid-cycle amendment is not documented. Sources: kodo:dash, kodo:advisor.
Restart persistence [supported] goal-plan.json and log.jsonl persist under ~/.kodo/runs/<id>/ and --resume reloads them. Sources: kodo:log, kodo:main.
Stages are an ordered list with optional parallel_group; no epic hierarchy, blockers or readiness model. Stages run sequentially unless grouped, and groups run in git worktrees with a persist_changes merge-back.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
Sequential/parallel grouping only; no dependency graph, epics or task growth.
Blockers [unsupported] No blocker field; ordering is by stage index. Sources: kodo:types.
Readiness [unsupported] No readiness state; the next stage starts after the previous finishes. Sources: kodo:types.
Enforcement [partial] Sequential order and parallel_group barriers are enforced by the orchestrator loop. Sources: kodo:base, kodo:parallel.
State propagation [partial] Stage results feed summaries to later stages and merge failures surface in the run result. Sources: kodo:base, kodo:gitops.
Task to epic [unsupported] Not supported: a goal cannot be grown into an epic that keeps identity. Sources: kodo:types.
Identity retention [unsupported] No task identity across runs; fix-from references a prior RUN_ID only. Sources: kodo:main.
Workflow constraints [supported] Goal, improve and test modes use fixed pipelines; goal mode intake plans stages automatically; teams are configurable but the staged loop is mandatory. Sources: kodo:readme, kodo:types.
Per-run summaries bridge cycles; the project can carry .kodo/config.json, .kodo/team.json and .kodo/test-coverage.md; agents read repository CLAUDE.md/AGENTS.md through their own CLIs. No cross-run semantic memory.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
Context is run-local summaries and files; no managed memory with citation, correction or deletion controls.
10 atomic answers
Instructions [partial] kodo passes role prompts and the goal; repository instruction files are read by Claude Code/Cursor/Codex natively and retained for OpenCode. Sources: kodo:roles, kodo:opencode.
Skills [unsupported] Bounded search found no skill mechanism in kodo. Sources: kodo:srch-api.
Project docs [partial] kodo writes .kodo/run-status.md and .kodo/test-coverage.md and reads goal.md. Sources: kodo:runstatus, kodo:roles.
Memory [partial] Cycle summaries carry context between cycles; kodo.knowledge targets non-code knowledge tasks; no cross-run memory store. Sources: kodo:readme, kodo:base.
Scope [partial] Per-run (summaries) and per-project (.kodo files). Sources: kodo:runstatus.
Provenance [partial] Events carry agent names and timestamps in log.jsonl; summaries are model-written. Sources: kodo:log.
Correction [unsupported] No UI or command to correct stored context. Sources: kodo:srch-api.
Deletion [partial] Run directories are plain files that can be deleted; no managed deletion. Sources: kodo:log.
Freshness [unsupported] No stale-context controls found. Sources: kodo:srch-api.
Takeover continuity [partial] A new session on the same machine can resume a run from logs; no multi-person takeover. Sources: kodo:main, kodo:log.
No code index or retrieval layer; agents search through their own CLIs. kodo adds a coach/advisor and an architect reviewer role but no architecture-conflict detection tied to a maintained model.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
Architecture awareness is prompt-driven; no indexing, citations or enforcement.
10 atomic answers
Code search [unsupported] No kodo-owned code search; delegated to the backend agents. Sources: kodo:srch-api.
Indexing [unsupported] No repository index found. Sources: kodo:srch-api.
External knowledge [partial] kodo.knowledge module handles research-style tasks through agents, with no retained external index. Sources: kodo:readme.
Multimodal inputs [partial] Claude sessions allow 16 MiB SDK messages for image-bearing tool results; no kodo image input flow. Sources: kodo:providers.
Freshness [unsupported] Not applicable to a missing index; no freshness metadata. Sources: kodo:srch-api.
Access scope [partial] Agents can access any file on the machine (README warning). Sources: kodo:readme.
Api retrieval [unsupported] No API retrieval endpoint. Sources: kodo:srch-api.
Citations [unsupported] No citation mechanism. Sources: kodo:srch-api.
Conflict detection [partial] The architect role reviews structure when asked; no automated conflict detection. Sources: kodo:roles, kodo:teamfull.
Review enforcement [unsupported] Review is requested by the orchestrator, not enforced. Sources: kodo:verif, kodo:readme.
No native issue management; work items are per-run stages, and `kodo issue` only opens a pre-filled GitHub issue for kodo itself.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
17 atomic answers
Create [unsupported] No issue entity; goals create runs and stages. Sources: kodo:srch-issues.
Edit [partial] Stage text editable via dashboard PUT only. Sources: kodo:dash.
Discuss [unsupported] No comments or discussion. Sources: kodo:srch-issues.
Prioritize [unsupported] No priority field. Sources: kodo:srch-issues.
Agents have no access to a shared durable issue store; stage plans and run files are private to the run.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
14 atomic answers
Shared ids [unsupported] No shared issue IDs. Sources: kodo:srch-issues.
Read [partial] Orchestrator can read .kodo/run-status.md that kodo writes. Sources: kodo:runstatus.
Create [unsupported] Agents do not create issues; the advisor can add stages. Sources: kodo:advisor.
Update [unsupported] No issue update tool exposed to agents. Sources: kodo:agenttools.
Claim [unsupported] No claim operation. Sources: kodo:srch-issues.
Assign [unsupported] Orchestrator dispatches to named roles, which is not issue assignment. Sources: kodo:agenttools.
State changes [partial] Orchestrator done() signals stage completion. Sources: kodo:verif.
Dependencies [unsupported] No dependency creation. Sources: kodo:types.
Children [unsupported] No child issue creation. Sources: kodo:types.
Results [partial] Agent reports return to the orchestrator and are logged. Sources: kodo:agenttools, kodo:log.
Human visibility [partial] Dashboard and viewer show logged agent events. Sources: kodo:dash, kodo:viewer.
Attribution [partial] Events record agent role names. Sources: kodo:log.
Permissions [unsupported] No permission model. Sources: kodo:srch-auth.
Conflicts [unsupported] No conflict handling for concurrent edits to plans. Sources: kodo:dash.
No living product specification: acceptance criteria are per-stage, and mission.md/docs in the kodo repo are its own maintainers documents.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
9 atomic answers
Persistent requirements [unsupported] Stage acceptance criteria last for one run; no cross-release requirements. Sources: kodo:types.
Human edits [partial] goal.md and stage criteria can be edited by hand. Sources: kodo:dash.
Agent edits [unsupported] No agent-maintained spec store. Sources: kodo:srch-api.
Revisions [unsupported] No revisions. Sources: kodo:dash.
Approvals [unsupported] No approval rules for specs. Sources: kodo:srch-api.
Implementation links [unsupported] No requirement-to-code links. Sources: kodo:srch-api.
Later retrieval [unsupported] No retrieval in later runs. Sources: kodo:srch-api.
Drift detection [unsupported] No drift detection. Sources: kodo:srch-api.
Review integration [partial] Verifiers can check per-stage acceptance criteria during done(). Sources: kodo:verif.
One run executes cycles sequentially; stages in the same parallel_group run concurrently in separate git worktrees, capped by KODO_MAX_PARALLEL (default 2). A terminal progress table, the dashboard and the log viewer show agents and stage status.
Scope
Surfaces: kodo CLI (interactive and non-interactive), kodo dashboard (local loopback web UI), kodo log viewer (HTML) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
Concurrency is per run and defaults to 2; there is no multi-run scheduler or documented quota beyond that variable.
5 atomic answers
Concurrent sessions [partial] Several kodo runs can be started separately; within a run only grouped stages run in parallel threads. Sources: kodo:base, kodo:parallel.
Concurrent tasks [supported] Stages with the same parallel_group execute concurrently via ThreadPoolExecutor, each in a worktree. Sources: kodo:base, kodo:parallel.
Quotas [partial] KODO_MAX_PARALLEL env var (default 2) bounds parallel stage workers; DEFAULT_MAX_EXCHANGES 30 and DEFAULT_MAX_CYCLES 5 bound a run. Sources: kodo:base, kodo:factory.
Observed concurrency [partial] Source defines the thread pool and per-thread event loops; this catalog does not run it, so actual overlap is not measured. Sources: kodo:parallel.
Status visibility [supported] Live terminal progress table, dashboard run list with events/stats, and viewer timeline show running, finished and failed work. Sources: kodo:log, kodo:dash, kodo:viewer.
The orchestrator assigns work to named team roles through dispatch tools; stages are assigned by the planner. There are no human assignees, queues, claims or routing rules.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
Assignment is internal orchestration, not user-facing ownership.
7 atomic answers
Human assignment [unsupported] No human assignment concept; the user supplies a goal. Sources: kodo:srch-issues.
Agent assignment [supported] The orchestrator dispatches directives to worker_fast, worker_smart, tester, tester_browser, architect. Sources: kodo:agenttools, kodo:teamfull.
Claims [unsupported] No claim mechanism. Sources: kodo:srch-issues.
Queues [partial] Stages are an ordered list consumed by the run loop. Sources: kodo:types.
Routing [partial] The orchestrator model decides which role gets a task; descriptions in team config guide it. Sources: kodo:teamfull.
Duplicate prevention [partial] Parallel stages get separate worktrees and branches; no duplicate-work detection across runs. Sources: kodo:gitops.
Product-owned orchestration: an orchestrator model (API, or Claude Code/Gemini/Codex/Cursor/Kimi CLI with an MCP bridge) delegates to team agents, each with its own backend, model and turn budget, and receives their reports. Verification roles are separate agents. An optional coach/advisor adds messages.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
12 atomic answers
Roles [supported] Orchestrator, worker_fast, worker_smart, tester, tester_browser, architect in the default team; custom roles via team.json. Sources: kodo:teamfull, kodo:teamcfg.
Model harness effort [supported] Per agent backend, model, max_turns, timeout_s, fallback_model; run-level --effort low/standard/high/max. Sources: kodo:teamfull, kodo:readme.
Messages [supported] Directives go to agents via dispatch tools; the coach pushes info/warning/correction messages to the orchestrator. Sources: kodo:agenttools, kodo:coach.
Results return [supported] Agent reports are returned to the orchestrator and logged. Sources: kodo:agenttools, kodo:log.
Cross harness [supported] Default team mixes Cursor and Claude agents under one orchestrator. Sources: kodo:teamfull.
Orchestration owner [supported] kodo supplies the orchestrator, tools and prompts; it is not a provider-native subagent feature. Sources: kodo:mcpserver, kodo:roles.
Activation [supported] Default on for every run; --coach enables the optional advisor; orchestrator chosen by --orchestrator. Sources: kodo:readme, kodo:params.
Authentication [supported] Orchestrator API key (GOOGLE_API_KEY/ANTHROPIC_API_KEY/OPENAI_API_KEY) or the CLI backend login. Sources: kodo:providers.
Live progress [supported] Terminal progress table and dashboard SSE stream show agent runs as they happen. Sources: kodo:log, kodo:dash.
Retained results [supported] log.jsonl records orchestrator tool calls and agent_run_end events; conversations are kept under the run directory. Sources: kodo:log, kodo:trace.
Followup [partial] Sessions persist for follow-up within a run (new_conversation flag resets); after completion, --fix-from RUN_ID starts a fix run from findings, and --resume continues unfinished runs. Sources: kodo:teamfull, kodo:main, kodo:readme.
Invocation [supported] Orchestrator invokes agents as tools; no manual agent invocation UI. Sources: kodo:agenttools.
No scheduler, event trigger or recurring task. README recommends external cron for overnight non-interactive runs, which is a manual workaround.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
8 atomic answers
Schedules [unsupported] Bounded search for cron/schedule found nothing; README only says non-interactive mode is for scripting and overnight cron jobs. Sources: kodo:srch-sched, kodo:readme.
Events [unsupported] No event triggers. Sources: kodo:srch-sched.
Recurring tasks [unsupported] No recurring tasks. Sources: kodo:srch-sched.
Completion conditions [partial] A run ends on done() acceptance, max cycles, or fatal agent errors. Sources: kodo:factory, kodo:verif.
Scheduler location [unsupported] No scheduler exists. Sources: kodo:srch-sched.
Overlaps [unsupported] No overlap handling for scheduled runs. Sources: kodo:srch-sched.
Retries [partial] Orchestrator API calls retry with backoff in the knowledge module and api.py; no run-level retry schedule. Sources: kodo:srch-sched.
Client independence [unsupported] No scheduler; a run lives in the foreground CLI process. Sources: kodo:srch-sched.
Stop by interrupting the CLI; resume with --resume. Steering is limited to appending lines to human_feedback.txt (with --coach); the dashboard feedback and stop endpoints are unimplemented stubs. No notifications.
Scope
Surfaces: kodo CLI (interactive and non-interactive), kodo dashboard (local loopback web UI) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
Pause, steer and stop controls are minimal; no notification or escalation channel.
10 atomic answers
Pause [unsupported] No pause command; interrupt then resume. Sources: kodo:srch-api.
Stop [partial] Ctrl-C interrupts (SIGINT cleanup is deferred in worktree teardown); dashboard stop POST only returns "stop requested" (TODO). Sources: kodo:dash.
Steer [partial] human_feedback.txt lines are queued to the orchestrator as human advice when the coach is active; dashboard feedback is a stub. Sources: kodo:coach, kodo:dash.
Notifications [unsupported] No notification channel. Sources: kodo:srch-notify.
Pending decisions [unsupported] No pending-decision queue; agents run bypassPermissions and the orchestrator decides. Sources: kodo:claudesess.
Escalation [partial] Fatal agent errors stop the run (FatalAgentError); no human escalation flow. Sources: kodo:agenttools.
Visible errors [supported] Errors are printed in the CLI and logged as events (e.g. auto_commit_error, coach_poll_error). Sources: kodo:log.
Chat terminal parity [unknown] kodo has no chat surface; the dashboard is read-only for steering, so parity is not applicable to a chat client. Sources: kodo:dash.
Retained history [supported] log.jsonl and conversations persist after a stop and are inspectable via kodo logs. Sources: kodo:log, kodo:viewer.
The dashboard and viewer show runs, stages, cycles, agent events and stats for one run at a time with a run list; no cross-epic fleet board.
Scope
Surfaces: kodo dashboard (local loopback web UI), kodo log viewer (HTML) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
Run list and per-run timeline only; no epics, dependency graph or attention queue.
8 atomic answers
Fleet fixture [partial] /api/runs lists recent runs with goal, orchestrator, model and finished state; no multi-epic board. Sources: kodo:dash.
Parent child [partial] Run > stage > cycle > agent events are navigable in the timeline. Sources: kodo:dashjs, kodo:viewer.
Dependencies [unsupported] No dependency visualisation. Sources: kodo:types.
Attention [unsupported] No attention or awaiting-input states. Sources: kodo:dash.
History [supported] Past runs and events remain browsable. Sources: kodo:dash, kodo:viewer.
Workload [partial] Live stats endpoint reports session costs and orchestrator cost; no workload view. Sources: kodo:dash.
Navigation errors [partial] Event log includes error events, but no dedicated failure view was found. Sources: kodo:dashjs.
Fixture mode [supported] Source-inspected only; --debug runs deterministic mocked backends that exercise the same UI paths. Sources: kodo:readme, kodo:dash.
Workflow is a fixed staged loop (goal, improve, test, fix-from) with configurable teams, effort and model per agent set in JSON or flags before the run; roles and relationships are visible in the live table and the retained log. There is no natural-language reconfiguration during execution.
Scope
Surfaces: kodo CLI (interactive and non-interactive), kodo dashboard (local loopback web UI) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
In-run changes are limited to human_feedback.txt advice and dashboard stage edits.
10 atomic answers
Quick task [partial] `--team quick` and --effort low give a lighter run; there is no single-task mode outside the staged loop. Sources: kodo:readme, kodo:teamcfg.
Growing epic [unsupported] A run cannot grow into an epic. Sources: kodo:types.
Presets [supported] Teams full, quick and test plus user teams; kodo teams add/edit/delete. Sources: kodo:readme, kodo:subcmd.
Natural language [partial] Goal intake interview refines the goal in natural language; team editing is interactive/JSON. Sources: kodo:readme, kodo:subcmd.
Runs wherever the CLI runs: your machine or any host or VPS you start it on. kodo has no managed cloud, hosted control plane or detach-and-reconnect service; a run lives in the foreground process and is resumed from logs.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: local, own-server · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
Laptop independence only by running kodo yourself on a remote host under tmux/nohup; kodo supplies no remote client.
12 atomic answers
Local [supported] Default mode: the CLI and agents run on the local machine. Sources: kodo:readme.
Byo remote [partial] Nothing prevents installing kodo on a VPS; no documented remote setup or remote client. Sources: kodo:srch-mobile, kodo:readme.
Self hosted [unsupported] No self-hosted service component; the dashboard is loopback-only. Sources: kodo:srch-mobile, kodo:dash.
Managed cloud [unsupported] No vendor cloud. Sources: kodo:srch-mobile.
Agents work in the shared project directory by default; parallel stage groups use separate git worktrees and branches that are merged back. No container or VM isolation; Claude runs with bypassPermissions and can access the whole machine.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
A worktree is a file-level separation, not a security sandbox; Codex uses workspace-write only.
12 atomic answers
Checkout [supported] Serial stages edit the project checkout directly. Sources: kodo:readme, kodo:base.
Branch [supported] Parallel stages use kodo-* branches merged with --no-ff. Sources: kodo:gitops.
Worktree [supported] create_worktree per parallel stage; stale worktrees (>6h) are cleaned. Sources: kodo:gitops, kodo:parallel.
Container [unsupported] No container option; docker is only detected for hints. Sources: kodo:srch-sandbox.
Os sandbox [partial] Codex backend runs with --sandbox workspace-write; Claude and others have no kodo sandbox. Sources: kodo:codexsess, kodo:claudesess.
Filesystem [partial] Worktrees separate parallel stage files; agents can still read any path on the host (README warning). Sources: kodo:readme, kodo:gitops.
Process [unsupported] Agents are host processes with no process boundary. Sources: kodo:srch-sandbox.
Network [unsupported] No network restriction. Sources: kodo:srch-sandbox.
Secrets [partial] ANTHROPIC_API_KEY is stripped from Claude worker environments; other environment variables are inherited. Sources: kodo:env.
Code separation [partial] Parallel branches keep code separate until merge; serial work shares one tree. Sources: kodo:gitops.
Security boundary [unsupported] README states agents can access any file; no hostile-code containment. Sources: kodo:readme, kodo:srch-sandbox.
Runs persist event logs and plans; `kodo --resume` finds incomplete runs and continues, resuming agent sessions where the backend supports it (Claude, Cursor, Codex, Gemini CLI, Kimi, Kiro). Crashes lose the in-flight agent turn; stale worktrees are cleaned.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
Recovery is manual (--resume); no automatic supervisor and no multi-day compute replacement story.
12 atomic answers
Client disconnect [partial] The CLI is the run process; closing it stops the run; --resume continues later. Sources: kodo:main.
Worker crash [partial] Fatal agent errors stop the run; fallback_model retries Claude with a smaller model; auto_commit failures are non-fatal. Sources: kodo:agenttools, kodo:teamfull, kodo:gitops.
Service restart [partial] There is no service; after a host reboot --resume restores from log.jsonl. Sources: kodo:log.
Expiry [unsupported] No sandbox expiry exists. Sources: kodo:srch-mobile.
Transcript reload [supported] kodo logs / viewer reload past runs from log.jsonl. Sources: kodo:viewer, kodo:log.
Process reconnect [unsupported] A new process cannot reattach to a dead one. Sources: kodo:main.
Task resume [supported] --resume reloads the goal plan and resumes backend sessions (Claude, Cursor chat ID, Codex thread, Gemini, Kimi, Kiro per docs). Sources: kodo:providers, kodo:main.
State reconstruction [supported] find_incomplete_runs rebuilds RunState from log events (cycle_end finished=false, no run_end). Sources: kodo:log.
Context loss [partial] Backends without session IDs (Kiro resumes last session per directory) may lose context; cycle summaries bridge the rest. Sources: kodo:providers.
Duplicate actions [unknown] Whether a resumed agent repeats an in-flight turn depends on each backend; not determinable from kodo source alone. Sources: kodo:main.
Manual repair [partial] Interrupted merges or worktrees may need manual git cleanup beyond the stale-worktree sweeper. Sources: kodo:gitops.
Multi day continuity [partial] Resume is cycle-granular and works across days on the same host; no replacement of compute. Sources: kodo:log, kodo:main.
kodo does not manage dependencies, caches, snapshots, services or ports: agents install what they need in the project. The test mode prompts testers to build their own fixtures and report blocked workflows.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
10 atomic answers
Dependencies [unsupported] No dependency management; testers install software themselves per prompt. Sources: kodo:roles.
Per-agent max_turns and timeout_s, adaptive idle timeout (300 s) with a 75 minute cap, per-cycle exchange and run cycle limits, Claude fallback model, and error-pattern detection for subscription/rate-limit failures. No token or spend caps.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
No enforced spend or token limits.
8 atomic answers
Time limits [supported] timeout_s per agent (e.g. 1800 s workers, 600 s architect); idle 300 s and hard 4500 s caps in agent.py. Sources: kodo:agent, kodo:teamfull.
Turn limits [supported] max_turns per agent; --exchanges per cycle (default 30); --cycles per run (default 5). Sources: kodo:teamfull, kodo:factory.
Token limits [unsupported] No token limit setting; bounded search found none. Sources: kodo:srch-budget.
Spend limits [unsupported] No spend limit; costs are tracked but not capped. Sources: kodo:srch-budget, kodo:log.
Timeouts [supported] Agent timeouts raise and are reported to the orchestrator. Sources: kodo:agent.
Retries [partial] fallback_model for Claude and HTTP retry/backoff in API orchestrator calls. Sources: kodo:teamfull, kodo:sesbase.
Provider failures [partial] Subscription/rate-limit/billing patterns are classified so unrecoverable errors stop work. Sources: kodo:sesbase, kodo:agenttools.
Enforcement [partial] Turn, exchange, cycle and timeout limits are enforced stops; no cost enforcement. Sources: kodo:factory, kodo:agent.
Structured log.jsonl events, conversations per agent, live terminal table, dashboard stats and viewer; token and cost per session with API versus virtual buckets. No success-rate statistics beyond what the log shows.
Scope
Surfaces: kodo CLI (interactive and non-interactive), kodo dashboard (local loopback web UI), kodo log viewer (HTML) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
Not separately cross-checked
Native Kimi and Kiro report no token usage; no multi-project statistics.
13 atomic answers
Logs [supported] ~/.kodo/runs/<id>/log.jsonl plus conversations/ with orchestrator and agent events. Sources: kodo:log.
Traces [partial] Event-level traces of tool calls and session queries; no OpenTelemetry-style spans. Sources: kodo:log.
Failure reasons [supported] run_end and error events record error type and message. Sources: kodo:log.
Usage export [partial] Run logs are plain JSONL for export; no dedicated export command. Sources: kodo:log.
Task stats [partial] Per-run summary with cycles and cost; no per-stage dashboard statistics beyond events. Sources: kodo:dash, kodo:log.
Agent stats [supported] Live stats endpoint returns per-session tokens and cost. Sources: kodo:dash, kodo:log.
Project stats [partial] `kodo runs ./project` lists runs per project; no aggregate view. Sources: kodo:subcmd.
Time stats [partial] Event timestamps and elapsed time in the viewer; no time-series analytics. Sources: kodo:viewer, kodo:log.
Failure status [supported] Finished and unfinished/errored runs are distinguished (run_end, error field). Sources: kodo:log.
Approval status [unsupported] No approvals exist to report. Sources: kodo:claudesess.
Reconnect consistency [partial] State is rebuilt from log events on resume; consistency is by log replay. Sources: kodo:log.
Performance method [unknown] README cites a 100-task SWE-bench comparison hosted by the maintainer; this catalog measures none and treats the figure as vendor-claimed (not catalogued as a benchmark). Sources: kodo:readme.
Responsiveness [unknown] No responsiveness claim or measurement found in the repo. Sources: kodo:readme.
Per-session and orchestrator cost in two buckets, API (real spend from the orchestrator API) and Virtual (subscription-covered cost that the Claude SDK reports), plus input/output tokens; aggregated per run. Native Kimi and Kiro do not report usage.
Scope
Surfaces: kodo CLI (interactive and non-interactive), kodo dashboard (local loopback web UI) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
Not separately cross-checked
No per-epic, alert, export or reconciliation features; Cursor and Gemini report zero cost.
14 atomic answers
Task cost [partial] Run totals and per-cycle summaries show cost; no per-task field beyond sessions. Sources: kodo:log, kodo:dash.
Session cost [supported] RunStats tracks tokens and cost_usd per session and shows them in the progress table and /api/run/{id}/stats. Sources: kodo:log, kodo:dash.
Child cost [supported] Each agent session is tracked separately, so worker/architect costs are visible. Sources: kodo:log, kodo:dash.
Epic cost [unsupported] No epic entity; run totals only. Sources: kodo:types.
Harness aggregation [partial] Totals sum reported costs across backends, but Cursor and Gemini CLI orchestrators set 0.0 and Kimi/Kiro report nothing. Sources: kodo:readme.
Billed vs estimated [supported] API bucket is real spend; Virtual bucket is what a subscription would have cost, labelled as not charged. Sources: kodo:readme, kodo:log.
Tokens [supported] input_tokens and output_tokens recorded per session where the backend reports them. Sources: kodo:log.
Subscription allowance [partial] Subscription use is shown as Virtual cost, not as an allowance remaining. Sources: kodo:readme.
Provider reconciliation [unsupported] No reconciliation with provider bills. Sources: kodo:srch-budget.
Compute cost [unsupported] No compute cost; runs are local. Sources: kodo:srch-budget.
Refresh delay [partial] Cost updates as sessions finish; table refresh is throttled in the CLI. Sources: kodo:log.
Alerts [unsupported] No budget alerts. Sources: kodo:srch-budget.
Export [partial] Costs live in log.jsonl; no export command. Sources: kodo:log.
Drilldown [partial] Viewer and dashboard drill from run to events with per-session cost. Sources: kodo:viewer, kodo:dash.
Concurrency is bounded by KODO_MAX_PARALLEL (default 2) and per-agent turn/timeouts; no CPU/RAM awareness, queues or check deduplication.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Single-user local tool. The loopback dashboard is not shared; no comments, presence or multi-user prompting.
Scope
Surfaces: kodo CLI (interactive and non-interactive), kodo dashboard (local loopback web UI) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
11 atomic answers
Visibility [unsupported] Dashboard binds to 127.0.0.1 only; no shared visibility feature. Sources: kodo:dash, kodo:srch-auth.
Discussion [unsupported] No discussion. Sources: kodo:srch-auth.
Comments [unsupported] No comments. Sources: kodo:srch-auth.
Presence [unsupported] No presence. Sources: kodo:srch-auth.
Coprompting [unsupported] Only one operator; human_feedback.txt is a local file. Sources: kodo:coach.
Ordering [unsupported] No multi-user input ordering. Sources: kodo:coach.
Conflicts [unsupported] No conflict handling. Sources: kodo:srch-auth.
Attribution [partial] Coach queue tags human feedback as source="human"; no user identities. Sources: kodo:coach.
Permissions [unsupported] No permission model. Sources: kodo:srch-auth.
Plan restrictions [supported] Free MIT software; no plans. Sources: kodo:license.
Owner online [supported] The owner process is the run; it must stay running. Sources: kodo:main.
No organizations, roles, SSO or provisioning; access is whoever can run the CLI or reach the loopback dashboard.
Scope
Surfaces: kodo CLI (interactive and non-interactive), kodo dashboard (local loopback web UI) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
9 atomic answers
Organization roles [unsupported] None in code. Sources: kodo:srch-auth.
Sso [unsupported] No SSO/OIDC/SAML code found. Sources: kodo:srch-auth.
Provisioning [unsupported] No SCIM or provisioning. Sources: kodo:srch-auth.
Unauthorized access [partial] Dashboard is loopback-bound, which blocks remote hosts by default; no auth if another local user connects. Sources: kodo:dash, kodo:srch-auth.
Revocation [unsupported] No grants to revoke. Sources: kodo:srch-auth.
Plan scope [supported] No plans; free software. Sources: kodo:license.
No kodo approval system: Claude agents run in bypassPermissions mode and Codex in workspace-write; the README warns they can access any file. Safer behaviour is only whatever each backend CLI offers.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
14 atomic answers
Command [unsupported] No command approvals; bypassPermissions. Sources: kodo:claudesess.
Tool [unsupported] No tool approvals. Sources: kodo:claudesess.
Network [unsupported] No network approvals. Sources: kodo:srch-sandbox.
Secret [unsupported] No secret approvals. Sources: kodo:env.
Delivery [unsupported] Auto-commit runs without approval and never pushes. Sources: kodo:gitops.
Approver identity [unsupported] No approver concept. Sources: kodo:claudesess.
Lifetime [unsupported] No approval scopes. Sources: kodo:claudesess.
Allow [unsupported] No allow rules. Sources: kodo:claudesess.
Append-only log.jsonl with timestamped events per run (agents, tool calls, errors, cost) and saved conversations; local files, no user attribution, export, retention or integrity guarantees.
Scope
Surfaces: kodo CLI (interactive and non-interactive), kodo log viewer (HTML) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
Plain editable local files; no audit retention or integrity model.
11 atomic answers
Actions [supported] Orchestrator tool calls and agent runs are logged with timestamps. Sources: kodo:log.
Approvals [unsupported] No approvals to record. Sources: kodo:claudesess.
Edits [partial] Plan edits through the dashboard overwrite goal-plan.json without history. Sources: kodo:dash.
Discussion [unsupported] No discussion. Sources: kodo:srch-auth.
Decisions [partial] Orchestrator done() acceptance/rejection and reviewer reports are logged. Sources: kodo:log, kodo:verif.
Cross user persistence [unsupported] Single user. Sources: kodo:srch-auth.
Cross session persistence [supported] Runs persist under ~/.kodo/runs across sessions. Sources: kodo:log.
Export [partial] JSONL files are directly copyable; viewer can export HTML. Sources: kodo:log, kodo:viewer.
Mutability [supported] Files are ordinary local files that can be edited or deleted. Sources: kodo:log.
Retention [unsupported] No retention policy; files stay until deleted. Sources: kodo:log.
Integrity [unsupported] No integrity protection or the word audit. Sources: kodo:log.
Review is by reading logs/viewer and the resulting Git diff in your own tools; kodo has reviewer agents (architect, testers) but no diff or review UI, request-changes flow or keyboard/narrow-screen guidance.
Scope
Surfaces: kodo CLI (interactive and non-interactive), kodo log viewer (HTML) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
No review interface of its own.
13 atomic answers
Pending reviews [unsupported] No pending-review list. Sources: kodo:srch-pr.
Change understanding [partial] Run summaries and the viewer explain what agents did; the code diff stays in Git. Sources: kodo:viewer, kodo:log.
Evidence navigation [partial] Viewer navigates events and conversations. Sources: kodo:viewer.
Comments [unsupported] No review comments. Sources: kodo:srch-pr.
Request changes [partial] The orchestrator rejects work when the architect/tester fail it and re-dispatches; this is agent-to-agent, not human request-changes. Sources: kodo:verif.
Shipping guidance [unsupported] No next-shipping-action guidance; auto-commit never pushes. Sources: kodo:gitops.
Effort [unsupported] No review effort estimates. Sources: kodo:srch-pr.
Missed evidence [partial] README says to check test and review evidence before trusting success status. Sources: kodo:readme.
Keyboard [unsupported] No documented keyboard review support in the viewer. Sources: kodo:viewer.
Narrow screen [unsupported] No narrow-screen support documented. Sources: kodo:viewer.
Context switching [partial] Run list and event timeline keep past runs available. Sources: kodo:dash.
Roles [supported] Architect, tester and tester_browser roles are explicit and described in team config. Sources: kodo:teamfull.
Retained context [supported] Reviewer reports are retained in logs. Sources: kodo:log.
No colleague takeover: runs live on one host in one user directory; another person with the same files could resume a run with --resume, but there is no ownership transfer, attribution or shared runtime.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
Same-host resume only; no ownership model.
12 atomic answers
Ongoing takeover [partial] Anyone with access to ~/.kodo/runs and the project can run kodo --resume. Sources: kodo:main, kodo:log.
Same task [partial] Resume reuses the run and goal plan. Sources: kodo:main.
Session history [supported] log.jsonl and conversations retained. Sources: kodo:log.
Decisions [partial] Decisions exist only as logged orchestrator events. Sources: kodo:log.
Artifacts [partial] Run directory files (summaries, status) remain. Sources: kodo:log.
Permissions [unsupported] No permission transfer. Sources: kodo:srch-auth.
Runtime context [partial] Backend sessions resume by ID where supported. Sources: kodo:providers.
Owner disconnect [unsupported] If the owner process stops, the run stops. Sources: kodo:main.
Ownership [unsupported] No ownership transfer. Sources: kodo:srch-auth.
Attribution [unsupported] No user attribution. Sources: kodo:srch-auth.
Context loss [partial] Backends without IDs lose context; summaries bridge cycles. Sources: kodo:providers.
Harness transfer [partial] --resume can pick a different team (README/launch) so a different harness can continue; transcripts are not converted. Sources: kodo:main.
Retained outputs are run logs, conversations, summaries, reports (including tester and reviewer outputs) and a coverage file; code changes live in the working tree or branches. No screenshot/recording gallery or previews.
Scope
Surfaces: kodo CLI (interactive and non-interactive), kodo log viewer (HTML) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
No artifact gallery, previews or revision-linked artifacts.
13 atomic answers
Diffs [partial] Diffs are Git state in the project; agents may include them in reports. Sources: kodo:gitops.
Documents [partial] Run summaries, .kodo/run-status.md and test reports are saved. Sources: kodo:runstatus, kodo:roles.
Screenshots [unsupported] No screenshot capture feature (tester_browser uses a browser through Claude chrome flag). Sources: kodo:teamfull.
Recordings [unsupported] No recordings. Sources: kodo:srch-mobile.
Test outputs [partial] Tester reports are retained in logs; no structured test result store. Sources: kodo:log, kodo:roles.
Reviewer outputs [supported] Architect and tester reports are logged in full in conversations. Sources: kodo:log, kodo:verif.
Gallery [unsupported] No gallery. Sources: kodo:viewer.
Task links [partial] Artifacts sit in the run directory keyed by run ID. Sources: kodo:log.
Revision links [unsupported] No linking to a commit revision other than auto-commit messages. Sources: kodo:gitops.
Persistence [supported] Files persist after sessions end. Sources: kodo:log.
Child findings [supported] Reviewer findings from completed agents stay in conversations/log. Sources: kodo:log.
Prompt leakage [unknown] Whether agent prompts leak into retained outputs depends on model behaviour and was not tested; no redaction other than the opt-in trace upload cleaner. Sources: kodo:trace.
Verification is orchestrator-directed: done() asks testers, a browser tester and the architect to verify and parses an ALL CHECKS PASS signal; the README states there is no enforced independent gate. No built-in lint/type/CI gate and no spec conformance.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
Checks are requested by prompt, not enforced policy; failure does not block delivery in code.
11 atomic answers
Tests [partial] Tester agents run tests when asked; test mode builds user-level tests. Sources: kodo:verif, kodo:roles.
Lint [unsupported] No lint step in kodo. Sources: kodo:verif.
Types [unsupported] No type-check step. Sources: kodo:verif.
Ci [unsupported] No CI integration. Sources: kodo:srch-pr.
Independent review [partial] Architect reviewer agent can independently review; completion path relies on the orchestrator requesting it. Sources: kodo:readme, kodo:verif.
Spec checks [partial] Stage acceptance_criteria are passed to verifiers (CRITERIA_VERIFICATION_INSTRUCTIONS); no persistent spec. Sources: kodo:types, kodo:verif.
Required checks [partial] Verifiers list in team config marks testers/reviewers; README says the default completion tool does not enforce them. Sources: kodo:teamfull, kodo:readme.
Failure blocking [partial] A rejected done() sends the orchestrator back to fix; nothing blocks Git. Sources: kodo:verif.
Requirement conflicts [unsupported] No requirement-conflict check. Sources: kodo:verif.
Configured policy [partial] verification full/skip per stage and QuickCheck file-existence checks. Sources: kodo:types.
Local Git only: worktrees and branches for parallel stages, --no-ff merges with agent-assisted conflict resolution, and an agent-run auto-commit (never pushes). No PR creation, review comments or CI response.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
No remote, PR, CI or rebase workflow.
9 atomic answers
Branches [supported] kodo-* worktree branches for parallel stages. Sources: kodo:gitops.
No approve, merge-authority, release or deploy workflow beyond local merges of kodo-created stage branches.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
6 atomic answers
Approve [unsupported] No approval step. Sources: kodo:srch-pr.
Merge [partial] kodo merges its own worktree branches into the project branch automatically when persist_changes is set. Sources: kodo:gitops, kodo:types.
Release [unsupported] No release feature. Sources: kodo:srch-pr.
Deploy [unsupported] No deploy feature. Sources: kodo:srch-pr.
Exact revision [unsupported] No exact-revision checks. Sources: kodo:gitops.
Authority enforcement [unsupported] No authority model. Sources: kodo:srch-auth.
Delivery evidence is commits created by auto-commit and log events; no task-to-PR, build or deployment links, rollback or stale-approval invalidation.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
Commit-level only; no PR/build/deploy links.
9 atomic answers
Commit links [partial] Auto-commit start/done events are logged with the summary; no commit hash link in the plan. Sources: kodo:gitops, kodo:log.
Pr links [unsupported] No PRs. Sources: kodo:srch-pr.
Build links [unsupported] No builds. Sources: kodo:srch-pr.
Deployment links [unsupported] No deployments. Sources: kodo:srch-pr.
Delivery status [partial] run_end and stage results show success; delivery means local commit. Sources: kodo:log.
Retained evidence [supported] Logs and reports remain. Sources: kodo:log.
Rollback [unsupported] No rollback feature. Sources: kodo:gitops.
Revert [unsupported] No revert feature (Git is available manually). Sources: kodo:gitops.
Stale approval invalidation [unsupported] No approvals to invalidate. Sources: kodo:srch-pr.
No batch review queue: each run reports its own summary; kodo issue and the viewer do not group, filter or approve completed work.
Scope
Surfaces: kodo CLI (interactive and non-interactive), kodo log viewer (HTML) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Open source under MIT (LICENSE and pyproject); single Python repository (CLI, dashboard, viewer, benchmark harness). No hosted service or closed extension.
Scope
Surfaces: kodo CLI (interactive and non-interactive), kodo dashboard (local loopback web UI), kodo log viewer (HTML) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
6 atomic answers
Client license [supported] MIT for the whole package including the CLI and dashboard. Sources: kodo:license, kodo:pyproject.
Server license [supported] No server component beyond the local dashboard; same MIT license. Sources: kodo:license, kodo:dash.
Hosted extensions [supported] No hosted extensions; the benchmark site is a separate maintainer-run page linked from the README. Sources: kodo:readme.
Source visibility [supported] All code is public on GitHub (default branch dev). Sources: kodo:ghrepo.
Source revision [supported] Inspected commit 1933f9a84cc43b74dd1c45896ec257714bdf86c7 (2026-09-26); PyPI 0.5.1 and tag 0.5.1 are older. Sources: kodo:ghrepo, kodo:pypi.
Hosted parity [supported] No hosted product exists to compare. Sources: kodo:srch-mobile.
Fully local: install the package, bring your own backends and keys. Run data is plain JSON/JSONL/Markdown under ~/.kodo/runs; no vendor account, so nothing needs cancelling.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
8 atomic answers
Self hosting [supported] The whole product is a locally installed CLI. Sources: kodo:readme.
Vendor dependencies [supported] Needs model/provider accounts for the backends you choose; no kodo vendor service. Sources: kodo:providers.
Export [supported] Run files are plain JSONL/Markdown that can be copied. Sources: kodo:log.
Import [partial] --resume reads existing run directories; no import from other tools. Sources: kodo:main.
Formats [supported] JSONL, JSON, Markdown and an HTML viewer. Sources: kodo:log, kodo:viewer.
Cancellation [supported] No account or subscription with kodo. Sources: kodo:license.
Account independence [supported] Fully usable with no kodo account. Sources: kodo:readme.
History portability [partial] Run directories are portable files; resume expects the original project path. Sources: kodo:log, kodo:main.
Code, prompts and logs stay local and go to whichever model providers your backends and orchestrator use. kodo has no telemetry SDK; an opt-in KODO_TRACE_UPLOAD=1 uploads redacted run traces to a maintainer-owned GCS/Firestore project.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
Not separately cross-checked
Provider policies are separate and not assessed; retention, training and residency of provider data are not kodo facts.
15 atomic answers
Code data [supported] Code stays on the host; backend agents and the orchestrator model receive code context according to their providers. Sources: kodo:readme, kodo:providers.
Prompts [supported] Prompts go to the chosen orchestrator API and backend CLIs. Sources: kodo:models, kodo:providers.
Credentials [supported] Keys come from the environment or .env; ANTHROPIC_API_KEY is stripped from Claude workers; trace upload redacts secrets with detect-secrets. Sources: kodo:env, kodo:trace.
Logs [supported] Logs are local under ~/.kodo/runs unless trace upload is enabled. Sources: kodo:log, kodo:trace.
Artifacts [supported] Local run directory. Sources: kodo:log.
Retention [supported] No kodo-side retention; files persist until you delete them. Sources: kodo:log.
Deletion [supported] Delete run directories manually. Sources: kodo:log.
Training [unknown] Model-provider training policies are third-party facts outside this repository. Sources: kodo:providers.
Residency [unknown] Provider-dependent; kodo adds no hosted storage unless trace upload is on (GCS bucket kodo-bench). Sources: kodo:trace.
Subprocessors [unknown] kodo lists no subprocessors; provider subprocessors are third-party facts. Sources: kodo:trace.
Telemetry disclosure [supported] The trace_upload module docstring and testability.md disclose the opt-in upload (the README does not mention it); there is no default telemetry. Sources: kodo:trace, kodo:testability.
Telemetry defaults [supported] Off by default (KODO_TRACE_UPLOAD must be 1/true/yes). Sources: kodo:trace.
Telemetry controls [supported] Env var enables/disables trace upload; there is nothing else to disable. Sources: kodo:trace, kodo:srch-telemetry.
Consent prompt [partial] No first-run consent prompt found; the upload is opt-in by environment variable. Sources: kodo:trace.
Provider separation [supported] Orchestrator provider and each backend provider are separate; docs list them separately. Sources: kodo:providers.
Free and open source (MIT) with no kodo price; users pay their model providers and subscriptions (Claude Max/Pro, Cursor, ChatGPT, API tokens).
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
12 atomic answers
Currency [supported] No kodo price; costs are provider charges in the provider currency. Sources: kodo:license.
Interval [supported] No billing interval; MIT software. Sources: kodo:license.
Seats [supported] No seats. Sources: kodo:license.
Overages [supported] None from kodo; providers may bill overages. Sources: kodo:providers.
Taxes [supported] No kodo billing, so no tax treatment. Sources: kodo:license.
Model cost [supported] Separate: API orchestrator tokens (README says about $0.13 per run for Gemini Flash) and backend subscriptions. Sources: kodo:readme, kodo:providers.
Compute cost [supported] Runs on your own machine. Sources: kodo:readme.
No catalog scenario pricing: kodo has no platform fee and the README gives one example (Gemini Flash orchestrator about $0.13 per run, workers on subscription). Team and compute scenarios are not modelled because workloads are not defined.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
Only a vendor example; no five-developer or compute estimates.
11 atomic answers
One person [partial] kodo fee 0; add one agent subscription plus orchestrator tokens (README example ~$0.13/run). Sources: kodo:readme.
Two people [partial] Two installs, two sets of provider accounts; kodo fee 0. Sources: kodo:license.
Five people [partial] Five installs; kodo fee 0; provider costs scale with each account. Sources: kodo:license.
Task assumptions [unknown] No catalog task assumptions defined for this product. Sources: kodo:readme.
Concurrency [partial] KODO_MAX_PARALLEL default 2 would affect token burn; no cost model. Sources: kodo:base.
Tokens [unknown] Token volumes are workload-dependent and not published by the vendor beyond one example. Sources: kodo:readme.
Compute hours [partial] Local compute: no vendor compute fee. Sources: kodo:readme.
Fixed cost [supported] Fixed kodo cost is zero. Sources: kodo:license.
Variable cost [partial] Variable: orchestrator API tokens and subscription usage. Sources: kodo:readme, kodo:providers.
Included cost [partial] Claude/Cursor/ChatGPT subscription usage is covered by those plans, not by kodo. Sources: kodo:providers.
Unknown cost [supported] Provider token and subscription prices are unknown to this catalog for this scenario. Sources: kodo:providers.
Pre-1.0 open-source project (Beta classifier, PyPI 0.5.1 on 2026-06-23); default branch active through 2026-09-26 with unreleased changes; support through GitHub issues (including an automated CI-assistant workflow). No SLA or status page.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
Release cadence is irregular and recent work is unreleased; no response commitments.
8 atomic answers
Release stage [supported] Development Status :: 4 - Beta on PyPI; tag 0.5.1. Sources: kodo:pyproject, kodo:pypi.
Access restrictions [supported] Public repo and package; no access restrictions. Sources: kodo:ghrepo.
Support [partial] GitHub Issues and a `kodo issue` command to file them; no other channel found. Sources: kodo:ghrepo, kodo:srch-issues.
Release activity [partial] PyPI 0.5.0 (2026-05-19) and 0.5.1 (2026-06-23); default-branch commit 2026-09-26 not released; no GitHub releases. Sources: kodo:pypi, kodo:ghrepo.
Status history [unsupported] No status page for a local tool. Sources: kodo:ghrepo.
Deprecation [unsupported] Bounded search of README, docs and repo root found no deprecation policy or notices (no CHANGELOG). Sources: kodo:readme, kodo:ghrepo.
Export options [supported] Local files remain usable. Sources: kodo:log.
Free open-source tool (MIT); a zero-cost route exists with Gemini CLI plus a free Google API key or Ollama for the orchestrator. Everything else depends on the agent subscriptions you already hold.
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
10 atomic answers
Ongoing free [supported] MIT software with no paid tier. Sources: kodo:license.
Trial [supported] Not applicable; free forever software. Sources: kodo:license.
Credits trial [supported] No credits. Sources: kodo:srch-budget.
Self host cost [supported] Your machine, plus provider costs. Sources: kodo:readme.
Paid only [supported] Not paid-only; no kodo payment exists. Sources: kodo:license.
Early access [supported] Not gated. Sources: kodo:ghrepo.
Free scenarios [supported] docs/free-tier.md gives a zero-cost setup; two-person and remote scenarios need your own setup. Sources: kodo:freetier.
Card required [supported] No card for kodo; providers may require one. Sources: kodo:freetier.
Usage caps [supported] Caps come from providers (e.g. Gemini free tier). Sources: kodo:freetier.
Paid evaluation [supported] No payment needed to evaluate kodo itself. Sources: kodo:freetier.
No kodo pricing: free MIT software; model and subscription costs are third-party. Classification: no paid plan (transparent: nothing to charge).
Scope
Surfaces: kodo CLI (interactive and non-interactive) · Version: default-branch commit 1933f9a (2026-09-26, dev); PyPI kodo-agent 0.5.1 (2026-06-23) is older · Plan: Open source (MIT), local CLI; user pays model and subscription providers · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-17 (5 days old)
Agent cross-check
2026-11-17
9 atomic answers
Units [supported] No pricing units. Sources: kodo:license.
Included limits [supported] None. Sources: kodo:license.
Cancellation [supported] Nothing to cancel. Sources: kodo:license.
Sales only unknowns [supported] None; no sales path exists. Sources: kodo:license.
Transparency classification [supported] Transparent: free open-source software under MIT with no paid tier located. Sources: kodo:license, kodo:readme.
Overnight multi-agent coding orchestrator CLI (Python 3.13+, uv tool install kodo-agent); API or CLI-backend orchestrator delegates to installed Claude Code, Cursor, Codex, Gemini CLI, Kimi and Kiro; MIT; agents run bypassPermissions; verification is orchestrator-directed, not an enforced gate.
Local loopback web dashboard with run list, event log, SSE stream and live stats; stage plan editing via PUT; POST feedback and stop handlers are stubs (TODO comments) that only acknowledge.
Optional --coach advisor assesses progress and can push messages; a human can append lines (priority prefixes info/warning/correction) to run_dir/human_feedback.txt, which the coach polls.
Each run writes log.jsonl events under ~/.kodo/runs/<run_id>/; per-session token and cost stats with API and virtual buckets; find_incomplete_runs supports resume.
Per-backend install/auth notes: Claude Code (Max/Pro, strips ANTHROPIC_API_KEY), Cursor, Codex (ChatGPT plan or API key), Gemini CLI, Kimi (native ACP), Kiro; orchestrator can use OpenAI, Gemini, Claude or local Ollama.
Kind
source-inspected
Locator
per-backend sections; Orchestrator API keys; Local Ollama
API orchestrator aliases for Anthropic, Gemini, OpenAI and others via pydantic-ai, plus ollama: models against localhost:11434; per-provider API-key env var checks.
Kind
source-inspected
Locator
OLLAMA_LOCAL, ensure_ollama_base_url, alias table, env_vars
kodo builds an internal MCP server exposing each team agent as a tool to a CLI orchestrator; it is a per-run stdio/SSE bridge, not a documented external API.
Creates git worktrees for parallel stages, merges branches with --no-ff, asks an agent to resolve merge conflicts, and dispatches a worker to commit after verification (never pushes).
done() asks tester/browser/reviewer agents to verify and parses an ALL CHECKS PASS or MINOR ISSUES FIXED signal from their text; README says this is orchestrator-directed rather than an enforced gate.
States trace upload is disabled unless KODO_TRACE_UPLOAD=1 and, when enabled, writes run archives to GCS and Firestore; a contributor document, not the README.