50 of 51 criteria have a sourced finding; 37 were also cross-checked by a second agent. Every comparison that includes MartinLoop 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 (`martin-loop`/`martin`), a local stdio/HTTP MCP server and a TypeScript SDK; release validated on Windows, macOS and Linux. No desktop, mobile or editor app; a closed hosted dashboard exists on the Pro plan.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), @martinloop/mcp server (stdio and HTTP transports), Hosted dashboard / Atlas (closed, Pro plan) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
desktop_os: No desktop app; the Node CLI is validated by hosted Windows, macOS and Linux release matrices.
mobile_apps: No mobile app in the repository or product pages; bounded search found no iOS/Android code.
browser: The closed hosted Pro dashboard (martinloop.com) shows runs, receipts and cost history; the open-source core has no browser UI and the hosted client cannot be inspected.
mobile_browser: Responsive behavior of the closed hosted dashboard is not documented; the open-source core has no browser UI.
editor_extensions: No own editor extension; the MCP server installs into VS Code/Cursor/Codex/Claude/Gemini hosts and a Claude plugin manifest ships.
task_actions: Start (`run`), review (`dossier`, `share`), resume (`resume`/`continue`) and, over local stdio MCP, pause/continue/cancel; approvals are flags set before a run and there is no live steering input.
notifications: No notification code in the pinned packages; terminal output and an optional Arcade game only.
cross_device_continuity: Runs persist in the local runs root (~/.martin/runs); optional hosted sync can upload receipts for viewing on another device but does not move execution.
9 atomic answers
Desktop os [partial] No desktop app; the Node CLI is validated by hosted Windows, macOS and Linux release matrices. Sources: martinloop:release-ci, martinloop:readme.
Mobile apps [unsupported] No mobile app in the repository or product pages; bounded search found no iOS/Android code. Sources: martinloop:srch-clients, martinloop:readme.
Browser [partial] The closed hosted Pro dashboard (martinloop.com) shows runs, receipts and cost history; the open-source core has no browser UI and the hosted client cannot be inspected. Sources: martinloop:pricing, martinloop:llms.
Mobile browser [unknown] Responsive behavior of the closed hosted dashboard is not documented; the open-source core has no browser UI. Sources: martinloop:pricing.
Editor extensions [unsupported] No own editor extension; the MCP server installs into VS Code/Cursor/Codex/Claude/Gemini hosts and a Claude plugin manifest ships. Sources: martinloop:readme, martinloop:plugin, martinloop:srch-clients.
Task actions [partial] Start (`run`), review (`dossier`, `share`), resume (`resume`/`continue`) and, over local stdio MCP, pause/continue/cancel; approvals are flags set before a run and there is no live steering input. Sources: martinloop:cli-ref, martinloop:mcp-tools, martinloop:leash.
Notifications [unsupported] No notification code in the pinned packages; terminal output and an optional Arcade game only. Sources: martinloop:srch-notify, martinloop:readme.
Cross device continuity [partial] Runs persist in the local runs root (~/.martin/runs); optional hosted sync can upload receipts for viewing on another device but does not move execution. Sources: martinloop:readme, martinloop:cli-ref, martinloop:hosted-handoff.
Native adapters for Codex CLI, Claude Code CLI and Gemini CLI, plus an OpenAI-compatible HTTP adapter for any compatible endpoint; the product controls sessions through those adapters and has an ordered provider fallback chain.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), martin-loop TypeScript SDK and adapters · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
session_integration: Native adapters launch the agent CLIs and observe workspace changes; the openai adapter embeds a constrained structured-edit loop; no session API control beyond process launch.
mixed_harness_operation: An explicit ordered `--fallback-engine` chain can move a run to another engine after provider failure; swarm agents mix is not documented for different harnesses in one run.
5 atomic answers
Harnesses [supported] Native CLI adapters for Codex, Claude Code and Gemini CLI (`--engine codex|claude|gemini|auto`). Sources: martinloop:models, martinloop:readme.
Model providers [supported] `--model` passes through unchanged; the `openai` engine reaches OpenAI, OpenRouter, Together, Fireworks, DeepSeek, Kimi, Qwen, Mistral and local Ollama/LM Studio/llama.cpp endpoints. Sources: martinloop:models.
Custom adapters [supported] SDK exports createClaudeCliAdapter, createCodexCliAdapter, createGeminiCliAdapter, createDirectProviderAdapter and createOpenAiCompatibleAdapter; an OpenCode-style adapter example is shipped. Sources: martinloop:readme, martinloop:examples.
Session integration [partial] Native adapters launch the agent CLIs and observe workspace changes; the openai adapter embeds a constrained structured-edit loop; no session API control beyond process launch. Sources: martinloop:models.
Mixed harness operation [partial] An explicit ordered `--fallback-engine` chain can move a run to another engine after provider failure; swarm agents mix is not documented for different harnesses in one run. Sources: martinloop:changelog, martinloop:models.
Uses the host agent CLI's own login (Claude Code, Codex, Gemini) or API keys for the OpenAI-compatible engine; local models work through Ollama/LM Studio; no vendor credits or MartinLoop-held accounts.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
vendor_credits: No MartinLoop-provided model credits; spend is billed by the provider.
account_switching: Engine is chosen per run with `--engine` and saved with `enable --engine`; no multi-account switcher within one vendor.
provider_restrictions: Provider terms on automated use of subscriptions are not stated in the repository; they are vendor policy.
7 atomic answers
Agent subscriptions [supported] Native CLI adapters run the authenticated host runtime, so an existing Claude Code, Codex or Gemini login is used; MartinLoop injects no hidden model. Sources: martinloop:readme, martinloop:models.
Api accounts [supported] The openai engine takes MARTIN_OPENAI_BASE_URL/API_KEY/MODEL for provider API accounts. Sources: martinloop:models.
Vendor credits [unsupported] No MartinLoop-provided model credits; spend is billed by the provider. Sources: martinloop:readme, martinloop:pricing.
Local models [supported] Local endpoints (Ollama, LM Studio, llama.cpp) via the openai engine; API key optional. Sources: martinloop:models.
Account switching [partial] Engine is chosen per run with `--engine` and saved with `enable --engine`; no multi-account switcher within one vendor. Sources: martinloop:readme, martinloop:cli-ref.
Vendor account requirement [supported] No MartinLoop account is needed for Core; hosted sync and the dashboard need MARTIN_API_TOKEN. Sources: martinloop:readme, martinloop:cli-ref.
Provider restrictions [unknown] Provider terms on automated use of subscriptions are not stated in the repository; they are vendor policy. Sources: martinloop:models.
Documented path: `npx martin-loop start`, `enable --engine`, `demo`, then `run "<task>" --verify "npm test"`; needs Node 20+ and a supported agent CLI. Works on a Git repo; no remote or cloud setup exists.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
no_git: Rollback and change observation rely on Git for the repository root; non-Git directory behavior is not documented (rollback code shells out to git).
hosted_git: `martin_create_pr` uses the GitHub CLI for PRs; no hosted-Git connection or app is managed by MartinLoop.
managed_cloud_setup: No managed-cloud execution exists; bounded search found no remote runner.
byo_vps_onboarding: No VPS installer or remote onboarding flow in the pinned release.
7 atomic answers
First task setup [supported] start → demo workspace → `run` with --verify, --budget-usd and --max-iterations; `run` auto-checks doctor, session-start and preflight. Sources: martinloop:readme, martinloop:cli-ref.
No git [partial] Rollback and change observation rely on Git for the repository root; non-Git directory behavior is not documented (rollback code shells out to git). Sources: martinloop:rollback, martinloop:secpol.
Local git without remote [supported] Runs, rollback boundaries and receipts are local; no remote is required. Sources: martinloop:rollback, martinloop:readme.
Hosted git [partial] `martin_create_pr` uses the GitHub CLI for PRs; no hosted-Git connection or app is managed by MartinLoop. Sources: martinloop:prtools.
Prerequisite guidance [supported] Node 20+, an installed and authenticated agent CLI, and a verifier command; `doctor` checks readiness. Sources: martinloop:readme, martinloop:cli-ref.
CLI with JSON output, a TypeScript SDK, and an MCP server exposing run, plan, status, pause/continue/cancel, dossier and PR tools over stdio (and an HTTP option); no REST API or webhooks in Core.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), @martinloop/mcp server (stdio and HTTP transports), martin-loop TypeScript SDK and adapters · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
api: No hosted REST API in Core; a hosted sync route exists only as a closed native control plane contract.
webhooks: No inbound or outbound webhook code in the pinned packages.
authentication: Local CLI/MCP use the OS user (a same-machine trust boundary); hosted sync uses MARTIN_API_TOKEN.
stability: Pre-1.0 (0.9.0); changelog states no CLI command or MCP tool removed in 0.9.0, no formal API stability policy found.
Api [partial] No hosted REST API in Core; a hosted sync route exists only as a closed native control plane contract. Sources: martinloop:hosted-handoff.
Authentication [partial] Local CLI/MCP use the OS user (a same-machine trust boundary); hosted sync uses MARTIN_API_TOKEN. Sources: martinloop:mcp-tools, martinloop:cli-ref.
Stability [partial] Pre-1.0 (0.9.0); changelog states no CLI command or MCP tool removed in 0.9.0, no formal API stability policy found. Sources: martinloop:changelog, martinloop:rel090.
MCP server is exposed to hosts (no MCP client stated), Claude plugin skill, and install helpers; no plugin marketplace or custom tool registry inside MartinLoop.
Scope
Surfaces: @martinloop/mcp server (stdio and HTTP transports), martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
Not separately cross-checked
mcp_client: MartinLoop exposes an MCP server for hosts; bounded search found no MCP client for external tools in the pinned packages (the host agent provides its own).
plugins: A Claude plugin manifest with a martinloop-govern skill ships; no MartinLoop plugin framework.
custom_tools: Custom adapters via the SDK; no user-defined tool registry for agents.
instructions_import: Reads AGENTS.md/CLAUDE.md-style repo context only through the host agent; MartinLoop writes `llms.txt` guidance for agents.
authentication: `martin mcp install/verify-install` write host config for codex, claude, gemini, cursor and vscode; OAuth is not part of it.
Instructions import [partial] Reads AGENTS.md/CLAUDE.md-style repo context only through the host agent; MartinLoop writes `llms.txt` guidance for agents. Sources: martinloop:readme.
Authentication [partial] `martin mcp install/verify-install` write host config for codex, claude, gemini, cursor and vscode; OAuth is not part of it. Sources: martinloop:readme, martinloop:cli-ref.
Node/TypeScript CLI, MCP server and SDK; no desktop or mobile client. The hosted dashboard is a closed web app whose implementation is not published.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), Hosted dashboard / Atlas (closed, Pro plan) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
desktop_implementation: No desktop client; a Node.js/TypeScript terminal CLI (package.json bin martin-loop).
mobile_implementation: No mobile client in the repository.
hybrids: No Electron/Tauri hybrid code found.
implementation_evidence: Package manifest and source show a pnpm TypeScript monorepo; the hosted dashboard (a React single-page app at martinloop.com) is closed.
4 atomic answers
Desktop implementation [unsupported] No desktop client; a Node.js/TypeScript terminal CLI (package.json bin martin-loop). Sources: martinloop:readme, martinloop:srch-clients.
Mobile implementation [unsupported] No mobile client in the repository. Sources: martinloop:srch-clients.
Hybrids [unsupported] No Electron/Tauri hybrid code found. Sources: martinloop:srch-clients.
Implementation evidence [partial] Package manifest and source show a pnpm TypeScript monorepo; the hosted dashboard (a React single-page app at martinloop.com) is closed. Sources: martinloop:readme, martinloop:pricing.
Intake is a prompt/objective given to `martin run` plus an optional swarm plan file; no issue import or external tracker integration.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
external_import: No Linear/Jira/GitHub issue import in the pinned packages.
chat_git_triggers: No chat or Git-event triggers; a CI example runs the CLI from GitHub Actions manually configured.
authoritative_tracker: No tracker; the JSONL run store is authoritative for runs only.
stable_ids: Runs have loop IDs and swarm IDs; there are no issue IDs.
sync: No tracker sync.
supported_edits: No issue edits exist.
external_roundtrips: No external tracker relationship.
9 atomic answers
Local creation [supported] `martin run "<objective>"` with acceptance criteria (`--accept`) and metadata creates a run record. Sources: martinloop:cli-ref.
Prompt intake [supported] Objective as positional argument or --objective; MCP martin_run/martin_plan. Sources: martinloop:cli-ref, martinloop:mcp-tools.
External import [unsupported] No Linear/Jira/GitHub issue import in the pinned packages. Sources: martinloop:srch-tracker.
Chat git triggers [unsupported] No chat or Git-event triggers; a CI example runs the CLI from GitHub Actions manually configured. Sources: martinloop:srch-sched, martinloop:examples.
A task contract carries objective, acceptance criteria, verifier plan, scope and budget; `preflight` and `martin_plan` preview it. No plan editor, revision history or mid-run amendment.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), @martinloop/mcp server (stdio and HTTP transports) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
Not separately cross-checked
plans: A Governed Run Plan (objective, verifier plan, scope, budget) is shown before work; it is a contract, not an editable plan document.
editors: No plan editor; flags, config file or SDK object define the contract.
revisions: No revision history for contracts; run records are append-only per run.
implementation_links: Receipts link verifier steps, changed files and rollback evidence to the run; no commit/PR link beyond PR summary.
amendment: Budget and limits stay bound to the original run; reconnecting does not reset them and there is no mid-run edit.
7 atomic answers
Plans [partial] A Governed Run Plan (objective, verifier plan, scope, budget) is shown before work; it is a contract, not an editable plan document. Sources: martinloop:readme.
Acceptance criteria [supported] `--accept <criterion>` (repeatable) adds acceptance criteria stored in the task contract. Sources: martinloop:cli-ref, martinloop:readme.
Implementation links [partial] Receipts link verifier steps, changed files and rollback evidence to the run; no commit/PR link beyond PR summary. Sources: martinloop:readme, martinloop:prtools.
Amendment [unsupported] Budget and limits stay bound to the original run; reconnecting does not reset them and there is no mid-run edit. Sources: martinloop:changelog, martinloop:rev-srch-edit.
Restart persistence [supported] Contract, verifier, scope, budget and evidence persist and `continue` reuses the same loop ID. Sources: martinloop:changelog.
Swarm plans carry tasks with dependsOn, assigned agents and write scopes; the scheduler starts tasks only when dependencies are verified. No epics, arbitrary hierarchy or task-to-epic growth.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
epics: No epic entity; a swarm has one parent contract and tasks.
hierarchy: Parent swarm → tasks → child runs and attempts; depth is fixed, not arbitrary.
state_propagation: Parent outcome comes only from the sealed parent receipt; child success cannot verify the parent.
task_to_epic: No conversion of a task into an epic.
identity_retention: No epic growth; task IDs persist within a swarm plan.
workflow_constraints: Swarm requires a plan file with explicit agent count and concurrency; single-run mode needs no preset.
Blockers [supported] Task `dependsOn` lists block readiness; recovery policy can waive dependencies only when allowed. Sources: martinloop:scheduler.
Readiness [supported] Scheduler selects queued/ready tasks whose dependencies are all satisfied. Sources: martinloop:scheduler.
Enforcement [supported] Dependency gating is enforced in the scheduler before dispatch, not just displayed. Sources: martinloop:scheduler.
State propagation [partial] Parent outcome comes only from the sealed parent receipt; child success cannot verify the parent. Sources: martinloop:hosted-handoff, martinloop:scheduler.
Task to epic [unsupported] No conversion of a task into an epic. Sources: martinloop:swarm-cli.
Identity retention [unsupported] No epic growth; task IDs persist within a swarm plan. Sources: martinloop:swarm-cli.
Workflow constraints [partial] Swarm requires a plan file with explicit agent count and concurrency; single-run mode needs no preset. Sources: martinloop:rel090, martinloop:swarm-cli.
Config file defaults, repo AGENTS.md/CLAUDE.md, a Claude plugin skill and an append-only JSONL memory store; no managed cross-harness memory with correction or deletion controls.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
instructions: `martin.config.yaml` defaults and generated llms.txt guidance; repo instruction files are read by the host agent.
skills: A martinloop-govern skill ships in the Claude plugin.
project_docs: Repo analyzer reads project files for scope and style signals.
memory: Append-only JSONL memory of consent, preference, pattern, feedback and budget entries under the runs root.
scope: Memory is per runs root (local user); no per-project/team scoping beyond workspace and project IDs.
provenance: Entries distinguish explicit user input from inferred patterns with confidence scores.
correction: Entries are never overwritten or edited by design.
deletion: Append-only; no delete command found in the memory store.
10 atomic answers
Instructions [partial] `martin.config.yaml` defaults and generated llms.txt guidance; repo instruction files are read by the host agent. Sources: martinloop:config, martinloop:readme.
Skills [partial] A martinloop-govern skill ships in the Claude plugin. Sources: martinloop:plugin.
Project docs [partial] Repo analyzer reads project files for scope and style signals. Sources: martinloop:readme.
Memory [partial] Append-only JSONL memory of consent, preference, pattern, feedback and budget entries under the runs root. Sources: martinloop:memory.
Scope [partial] Memory is per runs root (local user); no per-project/team scoping beyond workspace and project IDs. Sources: martinloop:memory.
Provenance [partial] Entries distinguish explicit user input from inferred patterns with confidence scores. Sources: martinloop:memory.
Correction [unsupported] Entries are never overwritten or edited by design. Sources: martinloop:memory.
Deletion [unsupported] Append-only; no delete command found in the memory store. Sources: martinloop:memory.
Freshness [partial] Inferred memories degrade by confidence; no staleness warnings in review. Sources: martinloop:memory.
Takeover continuity [partial] Run records and receipts survive a new session on the same machine; hosted sync can share receipts. Sources: martinloop:readme, martinloop:cli-ref.
No code index or knowledge retrieval; the product analyzes the repository for scope/style and risk maps and binds agent output to path boundaries. Architecture enforcement is limited to path and policy checks.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), @martinloop/mcp server (stdio and HTTP transports) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
Not separately cross-checked
code_search: No MartinLoop code search tool; the host agent searches code.
indexing: A repo analyzer and `martin://repo/risk-map` resource summarize risky paths; no persistent index.
external_knowledge: No external knowledge connectors in the pinned packages.
multimodal_inputs: Text objectives only; no image input handling found.
freshness: No index to go stale.
access_scope: Allow/deny path globs limit what the agent may write.
Conflict detection [partial] Scope, test-integrity and dependency checks detect disallowed changes, not architectural conflicts. Sources: martinloop:secpol.
Review enforcement [partial] Policy and verifier gates run after each attempt; no architecture-consistency review. Sources: martinloop:secpol, martinloop:readme.
Prioritize [partial] `triage` ranks saved runs by urgency and missing evidence; no human priority field. Sources: martinloop:cli-ref, martinloop:mcp-tools.
Label [partial] `--metadata key=value` attaches metadata to a run record. Sources: martinloop:cli-ref.
Metadata [partial] Run records hold objective, verifier, budget, scope, metadata, attempts and costs. Sources: martinloop:readme.
Durable states [supported] Lifecycle outcomes VERIFIED, STOPPED, NEEDS REVIEW and stop reasons persist in run records. Sources: martinloop:readme.
External tracker requirement [supported] An external tracker is required for any issue management; none is integrated. Sources: martinloop:srch-tracker.
An agent host can start, inspect, pause, cancel and summarize runs through MCP, but there are no shared issue IDs or fields because no issue tracker exists.
Scope
Surfaces: @martinloop/mcp server (stdio and HTTP transports) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
Not separately cross-checked
shared_ids: Loop and swarm IDs are shared between CLI, MCP and receipts; there are no issue IDs.
create: martin_run starts a governed run; no issue creation.
update: No tool edits an existing run record.
claim: No claim operation.
assign: No assignment operation.
state_changes: martin_pause, martin_continue and martin_cancel change run control state on a live local run.
dependencies: No tool creates dependencies; swarm dependsOn is set in a plan file.
children: No tool creates children outside swarm plans.
14 atomic answers
Shared ids [partial] Loop and swarm IDs are shared between CLI, MCP and receipts; there are no issue IDs. Sources: martinloop:mcp-tools.
Read [supported] martin_get_run, martin_list_runs, martin_get_attempt and resources read run records. Sources: martinloop:mcp-tools.
Create [partial] martin_run starts a governed run; no issue creation. Sources: martinloop:mcp-tools.
State changes [partial] martin_pause, martin_continue and martin_cancel change run control state on a live local run. Sources: martinloop:mcp-tools.
Dependencies [unsupported] No tool creates dependencies; swarm dependsOn is set in a plan file. Sources: martinloop:swarm-cli.
Children [unsupported] No tool creates children outside swarm plans. Sources: martinloop:swarm-cli.
Results [supported] martin_run_dossier, martin_dossier and martin_pr_summary return results and blockers. Sources: martinloop:mcp-tools.
Human visibility [partial] CLI dossier shows the same records; no attributed human view. Sources: martinloop:mcp-tools.
Attribution [partial] Control receipts record local_os_user/local_mcp_client principal, explicitly not a human identity. Sources: martinloop:mcp-tools.
Permissions [partial] Run control only over local stdio with a live control-capable executor; remote transports are rejected. Sources: martinloop:mcp-tools.
Conflicts [partial] Terminal-state precedence and replay protection govern control requests. Sources: martinloop:rel090.
Swarm mode runs multiple child agents with explicit max concurrency (default simulation 15 agents, demo concurrency 5); single runs are one attempt loop. Status via `swarm status --watch`.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
quotas: maxConcurrency is a positive integer set per swarm; no global quota or plan limit.
observed_concurrency: No concurrency measurements are published; the catalog does not measure it.
5 atomic answers
Concurrent sessions [supported] Swarm children run concurrently up to the explicit max concurrency; live runs need explicit agents and concurrency. Sources: martinloop:scheduler, martinloop:rel090.
Concurrent tasks [supported] Ready tasks start in parallel up to effective concurrency min(requested, parent). Sources: martinloop:scheduler.
Quotas [partial] maxConcurrency is a positive integer set per swarm; no global quota or plan limit. Sources: martinloop:scheduler.
Observed concurrency [unknown] No concurrency measurements are published; the catalog does not measure it. Sources: martinloop:readme.
Status visibility [supported] `swarm status --watch`, `swarm inspect` and `triage` show running, failed and finished work. Sources: martinloop:swarm-cli, martinloop:cli-ref.
Swarm plans assign tasks to agent IDs with write scopes and durable execution leases prevent concurrent provider children for a run; no humans, queues or routing UI.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
human_assignment: No human assignment.
claims: Execution leases give one run one provider child; no task claim by agents.
queues: Tasks sit queued/ready in a scheduler for one swarm; no cross-run queue.
routing: Recovery policy can reassign a task up to maxReassignmentsPerTask; routing.ts classifies manager vs direct per task.
Agent assignment [supported] Plan tasks carry assignedAgentId and writeScope. Sources: martinloop:swarm-cli.
Claims [partial] Execution leases give one run one provider child; no task claim by agents. Sources: martinloop:changelog.
Queues [partial] Tasks sit queued/ready in a scheduler for one swarm; no cross-run queue. Sources: martinloop:scheduler.
Routing [partial] Recovery policy can reassign a task up to maxReassignmentsPerTask; routing.ts classifies manager vs direct per task. Sources: martinloop:swarm-cli.
Duplicate prevention [supported] Durable execution leases prevent concurrent provider children for one run during reconnects. Sources: martinloop:changelog.
MartinLoop supervises worker agents it launches (child runs in a swarm, assigned agent roles) and returns sealed parent receipts; there is no coordinator agent talking to children.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
Not separately cross-checked
roles: Plan agents have role fields; parent contract, child runs, global verifier.
model_harness_effort: Engine/model per run; per-agent harness choice inside a plan is not documented; no effort setting.
messages: No inter-agent messaging.
cross_harness: Whether one swarm can mix harnesses is not stated in plan docs or checked code paths.
activation: `swarm plan/run` and `demo --swarm`; live mode needs explicit agents, concurrency and positive budget.
followup: Reassignment and recovery within policy; no open follow-up chat to a finished child.
12 atomic answers
Roles [partial] Plan agents have role fields; parent contract, child runs, global verifier. Sources: martinloop:swarm-cli.
Model harness effort [partial] Engine and model are chosen per run and fixed for every child of a swarm (parent-selected); there is no per-agent harness or model choice inside a plan and no effort setting. Sources: martinloop:models, martinloop:rev-live-engine.
Results return [supported] Child results roll into evidence, global verification and a sealed parent receipt. Sources: martinloop:hosted-handoff.
Cross harness [unsupported] One swarm uses a single parent-selected engine and model: the live runtime rejects a child whose adapter provider or model differs, so harnesses cannot be mixed inside one swarm. Sources: martinloop:rev-live-engine, martinloop:models.
Orchestration owner [supported] MartinLoop core owns orchestration; the coding agent still writes code. Sources: martinloop:readme.
Activation [partial] `swarm plan/run` and `demo --swarm`; live mode needs explicit agents, concurrency and positive budget. Sources: martinloop:rel090.
Authentication [supported] Uses the host agent login or API key for workers. Sources: martinloop:models.
Live progress [supported] `swarm status --watch` shows live child state. Sources: martinloop:swarm-cli.
Retained results [supported] Swarm dossier, verify and share operate on stored runs. Sources: martinloop:swarm-cli.
Followup [partial] Reassignment and recovery within policy; no open follow-up chat to a finished child. Sources: martinloop:swarm-cli.
Pause/continue/cancel through local stdio MCP with receipts; `swarm cancel` and `resume`/`continue` on the CLI; hard stops preserve history. No notifications or live steering.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), @martinloop/mcp server (stdio and HTTP transports) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
Not separately cross-checked
steer: No live message to the running agent; objective and limits are fixed per run.
notifications: No notification code.
pending_decisions: Blocked runs report the exact approval flag to rerun with; no queued decision inbox.
escalation: Policy failures stop or escalate to NEEDS REVIEW with a next safe action.
chat_terminal_parity: Control mutations work only over local stdio MCP and CLI; remote/HTTP transports reject them.
10 atomic answers
Pause [supported] martin_pause requests a pause at a safe boundary on a live local run. Sources: martinloop:mcp-tools.
Resume [supported] `resume`/`continue <loopId>` and martin_continue continue eligible runs with recorded context. Sources: martinloop:cli-ref, martinloop:changelog.
Pending decisions [partial] Blocked runs report the exact approval flag to rerun with; no queued decision inbox. Sources: martinloop:leash.
Escalation [partial] Policy failures stop or escalate to NEEDS REVIEW with a next safe action. Sources: martinloop:readme, martinloop:secpol.
Visible errors [supported] Failure class (13), stop reason and verifier output appear in the dossier. Sources: martinloop:readme.
Chat terminal parity [partial] Control mutations work only over local stdio MCP and CLI; remote/HTTP transports reject them. Sources: martinloop:mcp-tools.
Retained history [supported] Stopped and cancelled runs keep records and receipts. Sources: martinloop:rel090.
No fleet or epic UI in Core; `triage`, `runs list` and swarm status/inspect give terminal views, and a closed hosted dashboard lists loops.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), Hosted dashboard / Atlas (closed, Pro plan) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
fleet_fixture: Swarm demo shows 15 agents on a launch board in the terminal; the hosted fleet UI is closed.
parent_child: Swarm inspect shows parent, tasks, children and attempts.
dependencies: Task dependsOn is part of the swarm plan view.
attention: Triage ranks runs by urgency and missing evidence.
workload: Status shows running count against concurrency; no workload dashboard.
navigation_errors: Failure class and stop reason shown per run.
8 atomic answers
Fleet fixture [partial] Swarm demo shows 15 agents on a launch board in the terminal; the hosted fleet UI is closed. Sources: martinloop:readme, martinloop:pricing.
Parent child [partial] Swarm inspect shows parent, tasks, children and attempts. Sources: martinloop:swarm-cli.
Dependencies [partial] Task dependsOn is part of the swarm plan view. Sources: martinloop:scheduler.
Attention [partial] Triage ranks runs by urgency and missing evidence. Sources: martinloop:cli-ref.
History [supported] Run history and dossiers persist. Sources: martinloop:readme.
Workload [partial] Status shows running count against concurrency; no workload dashboard. Sources: martinloop:swarm-cli.
Navigation errors [partial] Failure class and stop reason shown per run. Sources: martinloop:readme.
Fixture mode [supported] `demo --swarm` is a deterministic provider-free simulation labelled non-governed. Sources: martinloop:readme.
Single governed run or swarm plan; no presets or natural-language configuration, and settings are fixed per run. Orchestration is visible through swarm inspect and receipts.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Runs execute on the user's machine (the agent CLI process on the local checkout). No BYO-remote, self-hosted service, managed cloud or hybrid execution; hosted sync only uploads receipts.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), Hosted dashboard / Atlas (closed, Pro plan) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
byo_remote: No remote host launcher in the pinned packages.
self_hosted: No self-hosted execution service; Enterprise VPC/on-prem is a sales offering whose code is not public.
managed_cloud: No managed-cloud runner; hosted Pro is a dashboard of receipts.
hybrid: Local execution with optional hosted receipt sync (`sync flush`).
provisioning: No compute provisioning.
resources: Uses host machine resources; no compute limits.
region: Hosted region/residency is not published; the hosted service is closed.
expiry: No remote sessions to expire; local runs persist in the runs root.
12 atomic answers
Local [supported] Governed runs execute locally with the agent CLI against the local repository. Sources: martinloop:readme, martinloop:secpol.
Byo remote [unsupported] No remote host launcher in the pinned packages. Sources: martinloop:srch-isolation.
Self hosted [unsupported] No self-hosted execution service; Enterprise VPC/on-prem is a sales offering whose code is not public. Sources: martinloop:pricing, martinloop:srch-isolation.
Region [unknown] Hosted region/residency is not published; the hosted service is closed. Sources: martinloop:pricing.
Expiry [unsupported] No remote sessions to expire; local runs persist in the runs root. Sources: martinloop:readme.
Client disconnect [partial] Closing the terminal ends the local process; `continue` can resume eligible interrupted reversible runs. Sources: martinloop:changelog.
Other device continuation [unsupported] Execution cannot move to another device. Sources: martinloop:hosted-handoff.
No sandbox for the agent process; MartinLoop observes repository changes with allow/deny path rules, secret-file rules and rollback. Swarm children use Git worktrees (or independent clones) at a baseline commit.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
branch: Compares changes with the commit checked out at attempt start; no branch creation.
container: No container launcher; bounded search found only docker patterns for command blocking.
vm: No VM isolation.
os_sandbox: Release notes state it is not an operating-system sandbox; Codex/Claude sandboxes belong to the provider.
filesystem: Allow/deny globs, symlink and junction collapse, secret-file protection and rollback check repo writes; writes outside the repo are not contained.
process: Agent process is not contained or limited.
network: `--allow-network-domain` with execution profiles governs declared network authority for staging; agent network access is the provider's.
secrets: Writes to .env/key files are blocked unless explicitly allowed; secret-like task text is refused.
12 atomic answers
Checkout [supported] Single governed runs act in the configured repository checkout (--cwd). Sources: martinloop:secpol, martinloop:cli-ref.
Branch [partial] Compares changes with the commit checked out at attempt start; no branch creation. Sources: martinloop:secpol.
Worktree [supported] Swarm child, integration and verifier workspaces are Git worktrees (or independent clones) at a baseline commit. Sources: martinloop:workspaces.
Container [unsupported] No container launcher; bounded search found only docker patterns for command blocking. Sources: martinloop:srch-isolation.
Os sandbox [unsupported] Release notes state it is not an operating-system sandbox; Codex/Claude sandboxes belong to the provider. Sources: martinloop:rel090, martinloop:secpol.
Filesystem [partial] Allow/deny globs, symlink and junction collapse, secret-file protection and rollback check repo writes; writes outside the repo are not contained. Sources: martinloop:secpol.
Process [unsupported] Agent process is not contained or limited. Sources: martinloop:secpol.
Network [partial] `--allow-network-domain` with execution profiles governs declared network authority for staging; agent network access is the provider's. Sources: martinloop:leash, martinloop:extout.
Secrets [partial] Writes to .env/key files are blocked unless explicitly allowed; secret-like task text is refused. Sources: martinloop:secpol.
Code separation [partial] Swarm children are separated in worktrees; single runs share one checkout. Sources: martinloop:workspaces.
Security boundary [unsupported] No hostile-code containment claimed; use an isolated sandbox for host-level write prevention. Sources: martinloop:rel090.
Durable JSONL records, execution leases, retained rollback boundaries and `continue` support reconnect after interruption for reversible runs; terminal/exhausted/irreversible runs fail closed rather than replay.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
client_disconnect: A closed terminal interrupts the local process; eligible interrupted runs can `continue`.
worker_crash: Workspace-observation failure ends the attempt as a typed failure with safe rollback; stale-lock recovery is hardened.
service_restart: Durable execution leases and control journals survive reconnects (PID reuse hardened).
expiry: No sandbox expiry exists.
process_reconnect: Reconnect reuses the same loop and one provider child via leases.
context_loss: Provider session context is not stated to be retained; MartinLoop retains its own evidence.
manual_repair: Quarantine recovery stays explicit when the workspace cannot be restored safely.
12 atomic answers
Client disconnect [partial] A closed terminal interrupts the local process; eligible interrupted runs can `continue`. Sources: martinloop:changelog.
Worker crash [partial] Workspace-observation failure ends the attempt as a typed failure with safe rollback; stale-lock recovery is hardened. Sources: martinloop:changelog, martinloop:rel090.
Service restart [partial] Durable execution leases and control journals survive reconnects (PID reuse hardened). Sources: martinloop:rel090.
Transcript reload [supported] `dossier`, `runs get` and `runs attempt` reload persisted records. Sources: martinloop:cli-ref.
Process reconnect [partial] Reconnect reuses the same loop and one provider child via leases. Sources: martinloop:changelog.
Task resume [supported] `continue <loopId>` reuses persisted task, verifier, scope, budget and evidence for eligible runs. Sources: martinloop:changelog.
State reconstruction [partial] Retained candidate rollback authority and latest verifier failure survive reconnects. Sources: martinloop:changelog.
Context loss [partial] Provider session context is not stated to be retained; MartinLoop retains its own evidence. Sources: martinloop:changelog.
Duplicate actions [supported] Terminal, exhausted, uncertain-active, irreversible and external-write runs are not replayed automatically. Sources: martinloop:changelog.
Manual repair [partial] Quarantine recovery stays explicit when the workspace cannot be restored safely. Sources: martinloop:rel090.
Multi day continuity [partial] Records persist locally and limits stay bound across reconnects; no compute replacement exists. Sources: martinloop:changelog.
No environment management: dependencies, caches, services and ports are the user's; rollback ignores dependency/build directories. Swarm worktrees are created from a baseline commit.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
dependencies: No dependency installation step; the demo asks the user to run npm install.
caches: Rollback skips file-by-file inspection of node_modules, dist and coverage directories.
snapshots: Rollback boundary snapshots of repository files; no environment snapshots.
services: No service management.
ports: No port allocation.
secrets: Secret-file write blocking only; no secret injection.
duplicate_installs: Swarm worktrees have no shared install handling in the pinned code.
Hard USD, token, iteration, provider-timeout and verifier-timeout limits enforced before the next attempt; soft limit warns; provider fallback is explicit.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
time_limits: Provider execution and verifier timeouts exist in request options; USD/iteration limits are the headline controls.
timeouts: verifyTimeoutMs and providerExecutionTimeoutMs are request fields.
provider_failures: Failures are classified into 13 runtime classes and may trigger the configured fallback engine; external-write recovery stays disabled.
8 atomic answers
Time limits [partial] Provider execution and verifier timeouts exist in request options; USD/iteration limits are the headline controls. Sources: martinloop:readme.
Spend limits [supported] `--budget-usd` hard cap and `--soft-limit-usd` warning; budget preflight blocks attempts that would exceed policy. Sources: martinloop:cli-ref, martinloop:readme.
Timeouts [partial] verifyTimeoutMs and providerExecutionTimeoutMs are request fields. Sources: martinloop:cli-ref.
Retries [supported] Bounded by max iterations; provider fallback chain is explicit and validated before spend. Sources: martinloop:changelog.
Provider failures [partial] Failures are classified into 13 runtime classes and may trigger the configured fallback engine; external-write recovery stays disabled. Sources: martinloop:readme, martinloop:changelog.
Enforcement [supported] Budget caps stop the next attempt (budget_exit); limits cannot reset on reconnect. Sources: martinloop:readme, martinloop:changelog.
JSONL run records, per-attempt receipts, verifier evidence, 13-class failure classification, HMAC integrity verdicts, triage and dossier views. No benchmark performance numbers are produced by this catalog.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), @martinloop/mcp server (stdio and HTTP transports) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
usage_export: `share` writes a redacted JSON/Markdown/SVG bundle; `audit --json` for Claude Code history; no CSV usage export.
task_stats: Per-run cost, attempts and verifier results in the dossier.
agent_stats: Runs record runtime/engine; no per-agent statistics view found.
project_stats: Records carry workspace/project IDs; project aggregation is a hosted feature (closed).
time_stats: Audit reports retry chains and cost over a day range; run durations in records.
approval_status: Approval flags and policy blocks appear in the failure reason.
performance_method: A benchmark workspace exists but the catalog does not run it and MartinLoop is not on ADEMark.
responsiveness: No responsiveness measurements published.
Usage export [partial] `share` writes a redacted JSON/Markdown/SVG bundle; `audit --json` for Claude Code history; no CSV usage export. Sources: martinloop:cli-ref, martinloop:readme.
Task stats [partial] Per-run cost, attempts and verifier results in the dossier. Sources: martinloop:readme.
Agent stats [partial] Runs record runtime/engine; no per-agent statistics view found. Sources: martinloop:readme.
Project stats [partial] Records carry workspace/project IDs; project aggregation is a hosted feature (closed). Sources: martinloop:cli-ref, martinloop:pricing.
Time stats [partial] Audit reports retry chains and cost over a day range; run durations in records. Sources: martinloop:readme.
Failure status [supported] Outcomes VERIFIED, STOPPED, NEEDS REVIEW with failure class. Sources: martinloop:readme.
Approval status [partial] Approval flags and policy blocks appear in the failure reason. Sources: martinloop:leash.
Reconnect consistency [supported] Terminal-state precedence and journals keep state consistent after reconnect. Sources: martinloop:rel090.
Performance method [unknown] A benchmark workspace exists but the catalog does not run it and MartinLoop is not on ADEMark. Sources: martinloop:readme.
Responsiveness [unknown] No responsiveness measurements published. Sources: martinloop:readme.
Per-run and per-attempt cost and tokens with provenance labels actual/calculated/estimated/unavailable; Codex usage is estimated unless the host exposes it; `audit` estimates API-equivalent spend in Claude Code history. No compute cost.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
Not separately cross-checked
child_cost: Swarm settlement reconciles child spend with the run ledger for live runs.
epic_cost: No epic entity; swarm total only.
harness_aggregation: Fallback chain keeps one run record; cross-harness aggregation beyond one run is not shown.
subscription_allowance: `audit` reports API-equivalent spend for subscription users; actual subscription marginal cost is not computed.
provider_reconciliation: Live Swarm requires settlement evidence that reconciles with the run ledger.
compute_cost: No compute fees exist.
refresh_delay: Usage refresh timing depends on the provider host and is not documented.
alerts: Soft limit threshold warns before the hard cap.
14 atomic answers
Task cost [supported] Each run has cost_usd with provenance in the proof receipt/dossier. Sources: martinloop:readme.
Session cost [supported] Run and attempt level cost in dossiers. Sources: martinloop:readme.
Child cost [partial] Swarm settlement reconciles child spend with the run ledger for live runs. Sources: martinloop:rel090.
Harness aggregation [partial] Fallback chain keeps one run record; cross-harness aggregation beyond one run is not shown. Sources: martinloop:changelog.
Billed vs estimated [supported] Provenance labels actual, calculated, estimated, unavailable are always shown. Sources: martinloop:readme.
Subscription allowance [partial] `audit` reports API-equivalent spend for subscription users; actual subscription marginal cost is not computed. Sources: martinloop:readme, martinloop:cli-ref.
Provider reconciliation [partial] Live Swarm requires settlement evidence that reconciles with the run ledger. Sources: martinloop:rel090.
Swarm scheduler enforces explicit maximum concurrency and dependency gating; sibling children get fair-share budget lease reservations. No CPU/RAM awareness, host awareness or check deduplication.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
Not separately cross-checked
cpu_ram: No CPU/RAM limits found in scheduler.
backpressure: Excess ready tasks wait for slots; no resource backpressure signal.
host_awareness: No per-host resource awareness.
check_reuse: No check reuse found.
check_deduplication: No check deduplication found.
bounded_tests: Verifier timeouts and external-outcome deadline with four concurrent observations.
fairness: Fair-share reservations are kept for sibling tasks.
recovery: Recovery policy reassigns tasks within maxReassignmentsPerTask.
12 atomic answers
Concurrency [supported] maxConcurrency must be a positive integer; effective = min(requested, parent maxConcurrency); live runs require it. Sources: martinloop:scheduler, martinloop:rel090.
Queues [supported] Tasks wait in queued/ready states until a slot and dependencies allow. Sources: martinloop:scheduler.
Cpu ram [unsupported] No CPU/RAM limits found in scheduler. Sources: martinloop:scheduler.
Backpressure [partial] Excess ready tasks wait for slots; no resource backpressure signal. Sources: martinloop:scheduler.
Host awareness [unsupported] No per-host resource awareness. Sources: martinloop:scheduler.
Fairness [partial] Budget leases for sibling children are reserved fair-share and reclaimed only as needed; no fairness policy for task or slot ordering is documented. Sources: martinloop:scheduler, martinloop:rev-sched-fair.
Recovery [partial] Recovery policy reassigns tasks within maxReassignmentsPerTask. Sources: martinloop:swarm-cli.
Shared limits [unsupported] One swarm has its own limits; no cross-swarm shared limits. Sources: martinloop:scheduler.
Single-user local tool; shared dashboards (Growth) are waitlist only. A share bundle can be handed to another person but there is no shared visibility, presence or co-prompting.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), Hosted dashboard / Atlas (closed, Pro plan) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
visibility: `share` produces a redacted receipt bundle for another person; the hosted Pro dashboard is personal-project only; Growth shared dashboards are waitlist.
discussion: No discussion code.
comments: No comments.
presence: No presence.
coprompting: No co-prompting; one local operator.
ordering: No multi-user input.
conflicts: No multi-user conflicts handled.
attribution: Control receipts record a local OS user principal only.
11 atomic answers
Visibility [partial] `share` produces a redacted receipt bundle for another person; the hosted Pro dashboard is personal-project only; Growth shared dashboards are waitlist. Sources: martinloop:readme, martinloop:pricing.
Membership [unknown] Membership lifecycle for the hosted product is not documented. Sources: martinloop:pricing.
Sso [unknown] Enterprise card lists SSO as contact-sales; not documented or testable. Sources: martinloop:pricing.
Provisioning [unknown] No provisioning documentation. Sources: martinloop:pricing.
Unauthorized access [partial] Local files and MCP control rely on the OS user; hosted sync needs a workspace token. Sources: martinloop:mcp-tools, martinloop:cli-ref.
Revocation [unknown] Hosted token revocation is not documented. Sources: martinloop:cli-ref.
Plan scope [partial] Enterprise only (contact sales); Free/Pro have no roles. Sources: martinloop:pricing.
Approvals are explicit pre-run flags (`--approve-test-changes`, dependency, migration, config, external writes) enforced by a policy leash with allow/deny path lists; no interactive approval queue.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
command: Verifier commands and destructive commands are checked and blocked; `destructiveActionPolicy: approval` in config.
tool: MCP tool profiles limit exposed tools; run control is stdio-only.
network: `--allow-network-domain` with execution profiles for staging runs.
delivery: `martin_create_pr` requires a VERIFIED run with verified receipt and `execute=true`.
approver_identity: Flags carry no approver identity; principal is the local OS user.
lifetime: Approvals apply to the single invocation that carries the flag.
revoke: Cancel a live run; no standing approvals to revoke.
tunnels: No tunnels.
14 atomic answers
Command [partial] Verifier commands and destructive commands are checked and blocked; `destructiveActionPolicy: approval` in config. Sources: martinloop:leash, martinloop:config.
Append-only run records with HMAC integrity (verified / tamper_detected / unsigned) preserve attempts, verifier results and control receipts; single-user, no export beyond share bundles, and the HMAC key is local.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
approvals: Approval flags used appear in the run contract/violations; no approver identity.
edits: Changed files and rollback evidence recorded per attempt.
discussion: No discussion.
decisions: Final outcome, stop reason and next safe action recorded.
cross_user_persistence: Single local user; hosted sync adds a workspace view only on paid plans.
mutability: Records are locally signed so edits are detectable, but a local user can alter files.
retention: Hosted Pro lists 30-day log retention; local retention is the user's.
integrity: HMAC per-runs-root key detects tampering; integrity is evidence, not a universal guarantee.
Review is terminal receipts: dossier, proof card, `martin_eval` grading and PR summaries/review helpers. No diff viewer or comment workflow in the product.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), @martinloop/mcp server (stdio and HTTP transports) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
change_understanding: Dossier lists changed files, verifier steps and cost; no diff UI.
evidence_navigation: `runs get/attempt/verify` navigate attempts and evidence.
comments: No review comments.
request_changes: No request-changes flow; `martin_review_pr` produces findings text.
effort: No reviewer effort measurements.
keyboard: Terminal CLI; no documented keyboard shortcuts.
narrow_screen: Not documented.
context_switching: `martin_pr_summary` and review helpers carry run evidence context.
13 atomic answers
Pending reviews [supported] `triage` and `martin_triage_runs` rank runs by urgency and missing evidence. Sources: martinloop:mcp-tools, martinloop:readme.
Change understanding [partial] Dossier lists changed files, verifier steps and cost; no diff UI. Sources: martinloop:readme.
Local single-user records; a second person can inspect a shared redacted receipt bundle but cannot take over a run, and hosted team workspaces are Growth waitlist.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), Hosted dashboard / Atlas (closed, Pro plan) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
ongoing_takeover: No takeover; runs execute under one local user.
same_task: `resume <loopId>` on the same machine reuses the same loop ID.
session_history: Records persist locally; share bundle exports a redacted summary.
decisions: Recorded in dossier.
artifacts: Receipts and share bundle files.
permissions: No shared permissions.
runtime_context: Runtime context (checkout, agent session) stays on the local host.
owner_disconnect: Local runs need the owner machine.
Retained artifacts are run records, receipts, verifier results, rollback evidence, diffs in the run store and proof-card SVG/Markdown/JSON share bundles; no screenshots, recordings or gallery.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
Not separately cross-checked
diffs: Repo-backed runs can persist diffs and rollback artifacts when a persistence store is configured.
documents: JSON and Markdown receipts under share/.
screenshots: No screenshot capture.
recordings: No recordings.
previews: No previews.
reviewer_outputs: `martin_review_pr` returns findings text; not stored as a retained reviewer artifact.
gallery: No artifact gallery.
revision_links: Receipts bind verifier evidence to the active run and workspace; commit linkage is in rollback boundaries.
13 atomic answers
Diffs [partial] Repo-backed runs can persist diffs and rollback artifacts when a persistence store is configured. Sources: martinloop:rollback.
Documents [partial] JSON and Markdown receipts under share/. Sources: martinloop:readme.
Task links [supported] Artifacts live under the run/loop. Sources: martinloop:readme.
Revision links [partial] Receipts bind verifier evidence to the active run and workspace; commit linkage is in rollback boundaries. Sources: martinloop:readme, martinloop:rollback.
Persistence [supported] Persist under the runs root after sessions end. Sources: martinloop:readme.
Completion requires fresh verifier evidence from a user-configured command (tests, lint, types or anything); failed verification blocks VERIFIED and PR creation; no spec gates.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), @martinloop/mcp server (stdio and HTTP transports) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
Not separately cross-checked
types: Any command; not built in.
ci: GitHub Actions budget-gate example; no CI result ingestion.
independent_review: Swarm global verifier is separate from child agents; `martin_review_pr` for PRs.
spec_checks: No spec conformance checks.
requirement_conflicts: No product requirement store.
Required checks [supported] A run cannot be VERIFIED without verifier evidence; test-removal needs approval. Sources: martinloop:readme, martinloop:secpol.
Failure blocking [supported] Failed verification yields no VERIFIED and PR creation refuses non-VERIFIED runs. Sources: martinloop:prtools, martinloop:readme.
Git is used for observation and rollback (reset --soft, snapshots) and swarm worktrees; PR creation only through `martin_create_pr` using gh. No commits, branches, review comments, CI response or rebase.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), @martinloop/mcp server (stdio and HTTP transports) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
branches: No branch creation; reads current branch for PR.
commits: Rewinds agent commits via git reset --soft when blocked; does not author commits itself.
comments: No PR comment posting found.
ci_response: No CI response handling.
rebase: No rebase code.
conflicts: Swarm integration workspace integrates child results; conflict handling beyond failure is not documented.
product_state: Run records keep verifier and rollback state; PR URL is returned, not stored.
Receipts link verifier evidence, changed files and rollback outcomes to the run; no commit/PR/build/deployment link store; rollback restores the workspace.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
commit_links: Rollback boundaries record the head commit at attempt start.
pr_links: PR creation returns a URL; not persisted in the run record.
build_links: No build links.
deployment_links: No deployment links.
delivery_status: VERIFIED/STOPPED/NEEDS REVIEW describe run outcome, not delivery.
revert: Restores workspace to pre-attempt state; no revert of merged PRs.
stale_approval_invalidation: Verifier evidence must be fresh and bound to the active run and workspace.
9 atomic answers
Commit links [partial] Rollback boundaries record the head commit at attempt start. Sources: martinloop:rollback.
Pr links [partial] PR creation returns a URL; not persisted in the run record. Sources: martinloop:prtools.
Delivery status [partial] VERIFIED/STOPPED/NEEDS REVIEW describe run outcome, not delivery. Sources: martinloop:readme.
Retained evidence [supported] Receipts and dossiers retain evidence. Sources: martinloop:readme.
Rollback [supported] Rollback boundary restore after failed/blocked attempts with explicit failed/unavailable states. Sources: martinloop:rollback, martinloop:secpol.
Revert [partial] Restores workspace to pre-attempt state; no revert of merged PRs. Sources: martinloop:rollback.
Stale approval invalidation [partial] Verifier evidence must be fresh and bound to the active run and workspace. Sources: martinloop:readme.
Core (CLI, MCP, SDK, adapters, plugin) is Apache-2.0 at v0.9.0; hosted product surfaces (dashboard, Atlas, control plane) are proprietary commercial and not in the repository.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), Hosted dashboard / Atlas (closed, Pro plan) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
hosted_parity: Hosted dashboard features are not in the repository; parity cannot be assumed.
Core is fully local with JSONL run store and no vendor account; hosted sync is optional. Enterprise VPC/on-prem is contact-sales. Exports are JSON/Markdown/SVG receipts.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), Hosted dashboard / Atlas (closed, Pro plan) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
self_hosting: Core runs locally; the hosted dashboard is closed; Enterprise VPC/on-prem is contact-sales.
import: No import command for run records found.
cancellation: Hosted Pro cancellation terms are in closed terms/billing pages not read.
history_portability: Run records are portable files; hosted history is not exportable per documentation read.
8 atomic answers
Self hosting [partial] Core runs locally; the hosted dashboard is closed; Enterprise VPC/on-prem is contact-sales. Sources: martinloop:readme, martinloop:pricing.
Vendor dependencies [supported] Core needs none beyond the chosen agent CLI/provider; telemetry and sync are optional. Sources: martinloop:readme.
Cancellation [unknown] Hosted Pro cancellation terms are in closed terms/billing pages not read. Sources: martinloop:pricing.
Account independence [supported] Core works without any account and with telemetry disabled. Sources: martinloop:readme.
History portability [partial] Run records are portable files; hosted history is not exportable per documentation read. Sources: martinloop:readme, martinloop:pricing.
Local-first: code and prompts stay local; hosted sync uploads preserved evidence only when configured. Telemetry is anonymous, allowlisted, on by default after a one-time disclosure and controllable. Providers see code per their own policy.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), Hosted dashboard / Atlas (closed, Pro plan) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
logs: Run logs stay local unless `sync flush` is configured.
artifacts: Share bundle is redacted and user-distributed.
retention: Hosted Pro: 30-day log retention; local files are the user's.
deletion: Hosted deletion process is not documented in read sources.
training: Hosted training-use policy is in closed privacy terms not read; provider policy applies to code sent to agents.
residency: No residency statement for the hosted service; telemetry endpoint is a Supabase function.
subprocessors: No subprocessor list read; telemetry names a Supabase endpoint.
consent_prompt: One-time non-blocking disclosure, not a consent prompt; blocked until shown.
15 atomic answers
Code data [supported] Source code, prompts and file paths are never sent in telemetry; agent providers receive code per their policies. Sources: martinloop:readme, martinloop:telemetry.
Credentials [supported] Secrets and env vars excluded from telemetry; tokens stay in environment; secret-like task text refused. Sources: martinloop:readme, martinloop:secpol.
Logs [partial] Run logs stay local unless `sync flush` is configured. Sources: martinloop:cli-ref.
Artifacts [partial] Share bundle is redacted and user-distributed. Sources: martinloop:readme.
Retention [partial] Hosted Pro: 30-day log retention; local files are the user's. Sources: martinloop:pricing.
Deletion [unknown] Hosted deletion process is not documented in read sources. Sources: martinloop:pricing.
Training [unknown] Hosted training-use policy is in closed privacy terms not read; provider policy applies to code sent to agents. Sources: martinloop:llms.
Residency [unknown] No residency statement for the hosted service; telemetry endpoint is a Supabase function. Sources: martinloop:readme.
Subprocessors [unknown] No subprocessor list read; telemetry names a Supabase endpoint. Sources: martinloop:readme.
Telemetry disclosure [supported] README lists what is and is not sent and the endpoint. Sources: martinloop:readme.
Telemetry defaults [supported] Enabled by default after a one-time disclosure on interactive installs. Sources: martinloop:readme, martinloop:telemetry.
Core is free (Apache-2.0). Hosted Pro is $49/month billed monthly (dashboard, run receipts, cost history, 30-day log retention, personal projects); Growth is waitlist-only and Enterprise is contact-sales. Model costs are separate.
Scope
Surfaces: Hosted dashboard / Atlas (closed, Pro plan), martin-loop CLI (npm package martin-loop, Node 20+) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core; hosted Pro $49/month · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
seats: Pro targets solo users with personal projects; per-seat pricing for Growth is not stated.
seat_minimum: Not stated for Growth/Enterprise.
workspace_fees: No workspace fee stated.
included_usage: Pro includes hosted dashboard, receipts, cost history and 30-day retention; no usage quotas are stated.
credits: No model credits.
trial: No trial stated on the pricing page text read.
overages: No overage terms stated.
taxes: Tax treatment not stated.
12 atomic answers
Currency [supported] USD on the pricing page. Sources: martinloop:pricing.
Interval [supported] Pro is billed monthly. Sources: martinloop:pricing.
Seats [partial] Pro targets solo users with personal projects; per-seat pricing for Growth is not stated. Sources: martinloop:pricing.
Seat minimum [unknown] Not stated for Growth/Enterprise. Sources: martinloop:pricing.
Workspace fees [unknown] No workspace fee stated. Sources: martinloop:pricing.
Included usage [partial] Pro includes hosted dashboard, receipts, cost history and 30-day retention; no usage quotas are stated. Sources: martinloop:pricing.
Core costs $0 for any number of developers; hosted Pro would be $49/month per user if each developer subscribed (personal projects only). Agent/model costs depend on the user's provider and are not modeled here.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), Hosted dashboard / Atlas (closed, Pro plan) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
one_person: Core $0; with Pro $49/month; model cost excluded.
two_people: Core $0 each; Pro is solo/personal projects and shared team dashboards are waitlist, so two collaborators cannot share one Pro workspace.
five_people: Core $0 each; five Pro subscriptions would be 5 × $49 = $245/month with no shared dashboard.
task_assumptions: No fixed workload is published; the catalog invents none.
tokens: Token spend is provider-specific.
variable_cost: Model/provider usage varies.
included_cost: Pro includes hosted dashboard features only.
unknown_cost: Growth/Enterprise prices are not published.
11 atomic answers
One person [partial] Core $0; with Pro $49/month; model cost excluded. Sources: martinloop:pricing.
Two people [partial] Core $0 each; Pro is solo/personal projects and shared team dashboards are waitlist, so two collaborators cannot share one Pro workspace. Sources: martinloop:pricing.
Five people [partial] Core $0 each; five Pro subscriptions would be 5 × $49 = $245/month with no shared dashboard. Sources: martinloop:pricing.
Task assumptions [unknown] No fixed workload is published; the catalog invents none. Sources: martinloop:readme.
Concurrency [supported] Swarm concurrency is set by the user; no pricing impact. Sources: martinloop:rel090.
Tokens [unknown] Token spend is provider-specific. Sources: martinloop:models.
Compute hours [supported] No compute-hours charge. Sources: martinloop:readme.
Fixed cost [supported] Pro $49/month fixed; Core $0. Sources: martinloop:pricing.
Pre-1.0 public releases every few days (v0.9.0 on 2026-11-03); support via GitHub issues and support@martinloop.com; no response commitments or status page found; hosted Growth is waitlist.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), Hosted dashboard / Atlas (closed, Pro plan) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
release_stage: Version 0.9.0 pre-1.0; hosted Pro generally available per pricing page, Growth waitlist.
access_restrictions: Growth waitlist and Enterprise contact-sales restrict access.
response_commitments: Enterprise lists an SLA but no terms are public.
status_history: No status page found.
deprecation: No deprecation policy found.
8 atomic answers
Release stage [partial] Version 0.9.0 pre-1.0; hosted Pro generally available per pricing page, Growth waitlist. Sources: martinloop:repoapi, martinloop:pricing.
Core is free and open source with no account; hosted Pro is paid ($49/month) with no stated trial; Growth is waitlist-only. Two-person collaboration, mobile and shared review are not free scenarios because they do not exist in Core.
Scope
Surfaces: martin-loop CLI (npm package martin-loop, Node 20+), Hosted dashboard / Atlas (closed, Pro plan) · Version: v0.9.0 (commit df38850, released 2026-11-03) · Plan: Open-source Core (Apache-2.0) · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Availability
generally-available
Delivery
native
Checked
2026-11-16 (6 days old)
Agent cross-check
2026-11-17
trial: No Pro trial stated.
credits_trial: No credits trial.
paid_only: Hosted dashboard is paid-only (Pro $49/month); Core is not.
early_access: Growth is waitlist only; Enterprise via sales.
free_scenarios: Free: local governed runs, BYO subscriptions, receipts and swarm demo; no free collaboration, remote compute, mobile or issue management.
card_required: Not stated for Core; Pro is checkout-based.
10 atomic answers
Ongoing free [supported] Open Source Core is free "forever" per the pricing page. Sources: martinloop:pricing.
Credits trial [unsupported] No credits trial. Sources: martinloop:pricing.
Self host cost [supported] Local execution; user pays providers only. Sources: martinloop:readme.
Paid only [partial] Hosted dashboard is paid-only (Pro $49/month); Core is not. Sources: martinloop:pricing.
Early access [partial] Growth is waitlist only; Enterprise via sales. Sources: martinloop:pricing.
Free scenarios [partial] Free: local governed runs, BYO subscriptions, receipts and swarm demo; no free collaboration, remote compute, mobile or issue management. Sources: martinloop:readme, martinloop:pricing.
Card required [unknown] Not stated for Core; Pro is checkout-based. Sources: martinloop:pricing.
Command list (start, doctor, run, preflight, dossier, resume, share, runs, sync, mcp) and run options for budget, iterations, tokens, engine, allow/deny paths and verifier.
Tools for plan, run, status, pause, cancel, continue, inspect, dossier, PR summary/create/review; run control only over local stdio with a live executor.
States supported agents Claude Code, Codex, Gemini CLI; Core is Apache-2.0 and hosted product surfaces are proprietary commercial; Atlas is a paid tenant-scoped surface.
No mobile or desktop app code; only a CLI, MCP server and SDK ship (matches are a CLI bridge that avoids the bundled Electron Node when hosted inside Claude Code desktop or VS Code, and license text).
Reviewer bounded search at pinned commit: 2 file hit(s), reviewed: packages/contracts/src/index.ts, plugins/martinloop/.codex-plugin/plugin.json. The two hits are an artifact-kind union member ('screenshot', contracts/src/index.ts line 274) and a plugin listing field (.codex-plugin/plugin.json line 33); neither is image input handling or capture code.
A live swarm plan carries a single engine and model; the runtime throws if a child adapter reports a different provider or model ("every child must use the parent-selected provider/model").
Swarm child, integration and verifier workspaces are Git worktrees (or independent clones) created at a baseline commit and removed after evidence persists.
The only "fair-share" wording concerns initial budget lease reservations for sibling children, reclaimed when a child needs more capacity; slot and task ordering carry no documented fairness policy.
Kind
source-inspected
Locator
scheduler.ts lines 150-200 (reserveSwarmBudgetLease, extendSwarmBudgetLease and its fair-share comment); lines 419-470 (concurrency limit)
Reviewer bounded search at pinned commit: 1 file hit(s), reviewed: packages/cli/src/audit.ts. The one hit is a verifier-command detection regex in audit.ts line 7 (mentions playwright test), not capture or preview code.
Anonymous allowlisted events after a one-time non-blocking disclosure; disabled by `martin telemetry off`, DO_NOT_TRACK, MARTIN_TELEMETRY_DISABLED or CI.