50 of 51 criteria have a sourced finding; 48 were also cross-checked by a second agent. Every comparison that includes Coder Agents 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.
Browser chat interface on the Coder dashboard Agents page plus an HTTP API and an experimental coder exp chat CLI; the coder CLI installs on Linux, macOS and Windows. There is no native desktop or mobile app for Agents and no Agents-specific editor extension; developers use their usual editor against the workspaces the agent produces.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface), Coder Agents HTTP API (/api/v2/chats) and coder CLI · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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 native mobile or desktop Agents client; phone-browser support is not stated.
Chat CLI is experimental.
9 atomic answers
Desktop os [unsupported] Agents has no desktop client; the control plane is self-hosted server software and the coder CLI runs on Linux, macOS and Windows. Sources: coder-agents:idx, coder-agents:desktop.
Mobile apps [unsupported] No iOS or Android Agents app in the docs agents section or repo apps. Sources: coder-agents:idx.
Mobile browser [partial] Agents page code has mobile viewport handling and responsive classes; docs do not state supported phone browsers. Sources: coder-agents:mobilecode.
Editor extensions [partial] Developers connect VS Code, JetBrains or Cursor to workspaces; no Agents chat extension found. Sources: coder-agents:ides, coder-agents:idx.
Task actions [supported] Start, steer (queued messages), approve plans, review diffs and resume chats in the dashboard. Sources: coder-agents:idx, coder-agents:gitprov.
Cross device continuity [supported] All state is in the Coder database; any device with dashboard access continues the chat. Sources: coder-agents:idx, coder-agents:arch.
Coder Agents embeds its own Go agent loop in the control plane and is not a wrapper for Claude Code or Codex; it can also delegate to external harnesses (Claude Code, Codex) as experimental ACP subagents running in the workspace. Model providers: Anthropic, OpenAI, Google, Azure OpenAI, Bedrock, OpenAI-compatible, OpenRouter, Vercel AI Gateway with custom base URLs.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface), Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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
ACP external harnesses are experimental and gated by an administrator flag.
5 atomic answers
Harnesses [partial] Own harness plus experimental ACP harnesses such as Claude Code and Codex. Sources: coder-agents:idx, coder-agents:acp.
Custom adapters [supported] OpenAI-compatible endpoints and ACP adapters configured as JSON in ~/.coder/acp. Sources: coder-agents:idx, coder-agents:acp.
Session integration [supported] Embedded loop in the control plane; ACP harnesses controlled over the Agent Client Protocol. Sources: coder-agents:idx, coder-agents:acp.
Mixed harness operation [partial] Coder Agents sub-agents plus ACP harness subagents in one chat (experimental). Sources: coder-agents:acp.
Admin-configured provider API keys (deployment-wide), optional per-provider user API keys (BYOK), model access through AI Gateway budgets, and local or self-hosted models through OpenAI-compatible endpoints. ACP harnesses use their own credentials, including whatever subscription login the harness supports. Vendor credits do not exist for the model itself; Agent Hours are purchased from Coder for AI Premium.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface), Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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 documented reuse of consumer agent subscriptions by the built-in agent.
7 atomic answers
Agent subscriptions [partial] ACP harnesses authenticate on their own inside the workspace; not documented as subscription reuse. Sources: coder-agents:acp.
Vendor credits [not-applicable] Coder does not resell model credits; AI Premium sells Agent Hours for agent runtime. Sources: coder-agents:lic.
Local models [supported] OpenAI-compatible self-hosted endpoints and custom base URLs. Sources: coder-agents:idx.
Account switching [partial] Personal model overrides when enabled; no multi-account manager. Sources: coder-agents:platform.
Vendor account requirement [supported] No vendor account for Community self-hosting; a license applies for Premium features. Sources: coder-agents:lic, coder-agents:pricing.
Provider restrictions [unknown] Third-party provider terms are not captured.
Requires a running Coder deployment, at least one template and an LLM provider key; admin configures providers and models in Admin settings > AI; users then open the Agents page and prompt; the agent selects a template and provisions a workspace automatically. Git and GitHub/GitLab external auth drive repo access; no local-directory mode.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface), Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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: Chats work without a workspace for Q&A; code tasks use workspaces whose repos come from templates.
local_git_without_remote: Workspaces can hold repos without remotes, but diff viewer and PR features need external auth.
managed_cloud_setup: There is no SaaS or managed Coder Agents component.
byo_vps_onboarding: Coder installs on user infrastructure with documented install paths; no VPS-specific installer for Agents.
7 atomic answers
First task setup [supported] Admin configures provider and model, then users chat. Sources: coder-agents:start.
No git [partial] Chats work without a workspace for Q&A; code tasks use workspaces whose repos come from templates. Sources: coder-agents:idx.
Local git without remote [partial] Workspaces can hold repos without remotes, but diff viewer and PR features need external auth. Sources: coder-agents:gitprov.
Hosted git [supported] GitHub and GitLab external auth for diffs and merge request status. Sources: coder-agents:gitprov.
Managed cloud setup [unsupported] There is no SaaS or managed Coder Agents component. Sources: coder-agents:idx.
Byo vps onboarding [partial] Coder installs on user infrastructure with documented install paths; no VPS-specific installer for Agents. Sources: coder-agents:desktop, coder-agents:start.
REST API under /api/v2/chats (list, create, watch, files, projects, model configs) with API keys; experimental coder exp chat CLI; Go SDK (codersdk); built-in MCP server (local stdio and remote HTTP with OAuth2) exposes Coder workspace tools to other assistants; CI and webhook triggering is expected through the API. The headless-agents API is listed under the AI Premium plan on the pricing page.
Scope
Surfaces: Coder Agents HTTP API (/api/v2/chats) and coder CLI, Coder dashboard Agents page (browser chat interface) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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 documented stability guarantee for chat endpoints; chat CLI is experimental; API plan gating unclear.
8 atomic answers
Cli [partial] coder exp chat is experimental. Sources: coder-agents:cli.
Api [supported] chat_routes defines authenticated REST endpoints. Sources: coder-agents:routes.
Sdk [supported] codersdk Go SDK includes chat types and feature flags. Sources: coder-agents:sdk.
Mcp server [supported] Built-in MCP server for Coder workspaces and templates. Sources: coder-agents:mcp.
Webhooks [partial] Docs suggest creating chats from GitHub webhooks via the API; no inbound webhook endpoint of its own. Sources: coder-agents:start.
Authentication [supported] API key middleware on chat routes. Sources: coder-agents:routes.
Stability [partial] Experimental markers on several features; no stability policy for chat API stated. Sources: coder-agents:acp, coder-agents:advisor.
MCP client through admin-approved external MCP servers and workspace .mcp.json, skills from workspace .agents/skills and ~/.coder/skills, instruction files (AGENTS.md, CLAUDE.md, .cursorrules), project instructions and memory, ACP harnesses and templates as extension points. Authentication for MCP servers and per-plan MCP access are controlled by admins.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface), Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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
plugins: Skills, MCP and ACP adapters; no plugin marketplace.
custom_tools: MCP and dynamic tools; no custom tool SDK documented.
Implementation evidence [supported] Repository contains site/ (web) and Go control plane; chat routes in coderd. Sources: coder-agents:routes, coder-agents:chatd.
Intake via prompts and attachments, API calls from CI and GitHub webhooks, and chat creation from other tools; no native issue import or two-way sync with external trackers.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface), Coder Agents HTTP API (/api/v2/chats) and coder CLI · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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 tracker integration documented.
9 atomic answers
Local creation [supported] Chats start from prompts in the UI. Sources: coder-agents:idx.
Prompt intake [supported] Prompts, attachments and API. Sources: coder-agents:idx.
External import [unsupported] No tracker import found in agents docs (searched issue, Linear, Jira). Sources: coder-agents:idx, coder-agents:start.
Chat git triggers [partial] Docs suggest GitHub webhook creation via the API as user-built integration. Sources: coder-agents:start.
Authoritative tracker [unsupported] No tracker concept; chats are the objects. Sources: coder-agents:tasks.
Stable ids [partial] Chat UUIDs; no issue IDs. Sources: coder-agents:routes.
Plan mode ("Plan first") lets the agent investigate and write a chat-specific plan file under .coder/plans, propose it for review with propose_plan and ask structured questions; plans persist in the chat and an Implement plan action resumes in normal mode. No acceptance-criteria field or revision history beyond the file.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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
acceptance_criteria: No acceptance-criteria feature found.
editors: Agent edits the plan file; user reviews proposed plan.
revisions: propose_plan snapshots the plan into the transcript.
implementation_links: Plan file lives in the workspace repo path.
7 atomic answers
Plans [supported] Plan mode with plan file. Sources: coder-agents:idx.
Acceptance criteria [unsupported] No acceptance-criteria feature found. Sources: coder-agents:idx.
Editors [partial] Agent edits the plan file; user reviews proposed plan. Sources: coder-agents:idx.
Revisions [partial] propose_plan snapshots the plan into the transcript. Sources: coder-agents:idx.
Implementation links [partial] Plan file lives in the workspace repo path. Sources: coder-agents:idx.
Amendment [supported] Queued messages can change direction during work. Sources: coder-agents:idx.
Admin system prompt, project instructions, skills and durable project memory (early access), workspace instruction files and skills snapshots pinned per chat with refresh, chat history in the database; per-user and per-org scopes; retention and auto-archive controls.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface), Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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
provenance: Changes recorded in audit log; no per-note citation.
correction: Memory editable in project page.
takeover_continuity: Chat sharing is read-only; project private to owner.
Code search and edits through workspace tools (read_file, execute, edit_files), web_search when enabled, MCP tools, image and PDF attachments, computer-use sub-agent with virtual desktop; no repository index, citations or architecture checks.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface), Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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
indexing: No index found in agents docs.
freshness: Live workspace reads; context snapshots refreshed on demand.
api_retrieval: API exposes chats and files.
citations: None.
10 atomic answers
Code search [supported] read_file and shell tools in the workspace. Sources: coder-agents:idx.
Indexing [unsupported] No index found in agents docs. Sources: coder-agents:idx.
External knowledge [supported] web_search and MCP tools. Sources: coder-agents:idx.
Multimodal inputs [supported] Image, PDF and text attachments and computer-use agent. Sources: coder-agents:idx.
Freshness [partial] Live workspace reads; context snapshots refreshed on demand. Sources: coder-agents:extend.
Access scope [supported] Tools run with the chat owner RBAC identity. Sources: coder-agents:idx.
Api retrieval [partial] API exposes chats and files. Sources: coder-agents:routes.
Chats, projects, search with filters (including pr_status), pinning, archive, auto-archive, sharing; there is no issue tracker: tasks were migrated to chats.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
No shared issue records; agents operate on chats, workspaces and sub-agents.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface), Coder Agents HTTP API (/api/v2/chats) and coder CLI · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Many chats run in parallel; Community and Premium licenses allow five concurrently active agents deployment-wide (queued beyond that); chatd defaults to 5 concurrent root agents and 10 concurrent subagents; AI Premium removes the cap against purchased Agent Hours. Sidebar lists chats with states.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface), Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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
observed_concurrency: Not measured here.
status_visibility: Chat list and watch stream; fleet board not described.
5 atomic answers
Concurrent sessions [supported] Multiple chats and workspaces per user. Sources: coder-agents:idx.
Concurrent tasks [supported] Sub-agents run in parallel. Sources: coder-agents:idx.
Quotas [supported] Five concurrent agents on Community and Premium; defaults 5 root and 10 sub-agents in code. Sources: coder-agents:lic, coder-agents:chatd.
Observed concurrency [unknown] Not measured here.
Status visibility [partial] Chat list and watch stream; fleet board not described. Sources: coder-agents:routes.
Work is queued when the concurrency pool is full; the root agent assigns work to sub-agents; no human assignment or claims.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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: None.
claims: None.
routing: Template selection by the agent from descriptions.
Product-owned delegation: root agents spawn general, explore and computer-use sub-agents in parallel, optionally on a chosen model, with list_agents, wait_agent, message_agent and interrupt_agent; sub-agents cannot create workspaces or spawn more. Experimental ACP subagents add external harnesses. Results return via wait_agent; sub-agent chats are viewable by link.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface), Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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
cross_harness: ACP harnesses experimental.
12 atomic answers
Roles [supported] General, explore, computer_use sub-agent types. Sources: coder-agents:idx.
Model harness effort [supported] Optional model_config_id per spawn; ACP harnesses. Sources: coder-agents:idx, coder-agents:acp.
Messages [supported] message_agent and interrupt_agent. Sources: coder-agents:idx.
Activation [supported] On by default; ACP and advisor need experiments; computer-use needs admin virtual desktop. Sources: coder-agents:idx, coder-agents:acp.
Authentication [supported] Sub-agents inherit the user identity; ACP uses harness credentials. Sources: coder-agents:arch, coder-agents:acp.
Live progress [supported] Sub-agent chats stream live. Sources: coder-agents:share.
Retained results [supported] Persisted in the database. Sources: coder-agents:arch.
Followup [supported] message_agent follow-ups; interrupt and resume. Sources: coder-agents:idx.
No scheduler in Coder Agents; API calls from CI or webhooks can start chats as user-built automation; auto-archive and retention jobs are maintenance only.
Scope
Surfaces: Coder Agents HTTP API (/api/v2/chats) and coder CLI, Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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
schedules: Bounded search of docs/ai-coder/agents found no schedule feature (hits only compaction and scheduling of pods).
events: Docs list CI and GitHub-webhook triggering through the API as a use case.
recurring_tasks: None.
completion_conditions: None.
8 atomic answers
Schedules [unsupported] Bounded search of docs/ai-coder/agents found no schedule feature (hits only compaction and scheduling of pods). Sources: coder-agents:idx, coder-agents:start.
Events [partial] Docs list CI and GitHub-webhook triggering through the API as a use case. Sources: coder-agents:start.
No mandatory workflow: describe the work, the agent chooses template and creates a workspace; Plan first mode, per-spawn model choice, admin-set system prompt, live sub-agent chats and retained transcripts.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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
growing_epic: No epic.
presets: Plan first mode; templates act as environment presets.
10 atomic answers
Quick task [supported] Describe work and run. Sources: coder-agents:idx.
The agent loop runs in the self-hosted control plane and tools run in Coder workspaces on infrastructure you operate (Kubernetes, VMs, cloud); chats and workspaces continue when the browser disconnects and can be continued from any device. There is no vendor cloud.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd), Coder dashboard Agents page (browser chat interface) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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 vendor-managed compute; you supply the Coder deployment and workspaces.
12 atomic answers
Local [partial] Workspaces can be local to the control-plane host depending on templates; not documented as a mode. Sources: coder-agents:idx.
Each agent works in a Coder workspace provisioned from an admin template (container, pod or VM depending on template); workspace isolation per user; optional template pattern puts the agent in a separate pod with network policy (early stage); no git worktree isolation described.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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
Default agent workspace has the same network access as a user workspace.
12 atomic answers
Checkout [supported] Repo in the workspace. Sources: coder-agents:idx.
Branch [partial] Branches via git in workspace. Sources: coder-agents:arch.
Worktree [unsupported] No worktree feature in agents docs. Sources: coder-agents:idx.
Container [supported] Kubernetes pod templates documented. Sources: coder-agents:isol.
Vm [partial] Templates may use VMs such as EC2; not agent-specific. Sources: coder-agents:ides.
Os sandbox [partial] Bubblewrap with firewall suggested; agent firewall docs exist. Sources: coder-agents:isol.
Filesystem [partial] Separate home directories in the two-pod pattern. Sources: coder-agents:isol.
Chat state persists in the Coder database so conversations survive workspace stop, delete or rebuild; the agent can resume in a new workspace from the last git branch; context compaction; queued messages; no automatic process re-attachment.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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
worker_crash: DB state persists; in-flight step lost.
service_restart: Persisted chats resume; details not enumerated.
expiry: Workspace expiry per template; chat survives.
process_reconnect: process_list tracks processes in a live workspace.
12 atomic answers
Client disconnect [supported] Loop runs in control plane. Sources: coder-agents:idx.
Worker crash [partial] DB state persists; in-flight step lost. Sources: coder-agents:arch.
Service restart [partial] Persisted chats resume; details not enumerated. Sources: coder-agents:arch.
Expiry [partial] Workspace expiry per template; chat survives. Sources: coder-agents:idx.
Process reconnect [partial] process_list tracks processes in a live workspace. Sources: coder-agents:idx.
Task resume [supported] New workspace from last known state. Sources: coder-agents:idx.
State reconstruction [partial] Git branch plus chat history. Sources: coder-agents:idx.
Context loss [partial] Compaction summarizes older context. Sources: coder-agents:arch.
Duplicate actions [partial] Receivers tolerate duplicate deliveries and state updates are deduplicated, but re-execution of interrupted tool calls is not addressed. Sources: coder-agents:archmd.
Manual repair [partial] /compact clears error states. Sources: coder-agents:arch.
Multi day continuity [supported] Persistent chats with 30-day archive retention. Sources: coder-agents:retention.
Environment setup is owned by Coder templates (Terraform), including dependencies, services and secrets; the agent only chooses a template and parameters; no agent-specific cache or port handling.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Deployment limits for steps per turn, generation retries, queued messages and attachments; exponential-backoff retry of classified transient provider errors; execute tool waits 10s by default; Agent Hours and concurrent-agent caps (5 on Community/Premium); AI Gateway per-user monthly budgets; context compaction thresholds and advisor limits.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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: execute waits 10s by default with a caller timeout; Agent Time pool; no whole-run time limit found.
token_limits: Advisor max output tokens; gateway budgets.
8 atomic answers
Time limits [partial] execute waits 10s by default with a caller timeout; Agent Time pool; no whole-run time limit found. Sources: coder-agents:execcode, coder-agents:lic.
Audit logs, chat debug logging and retention, AI Gateway monitoring and audit, Agent Time reporting to admins, in-app usage warnings, GitHub and PR status in chats; native per-chat cost pages were removed in favor of AI Gateway reports.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd), Coder dashboard Agents page (browser chat interface) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Spend per user and group through AI Gateway budgets and reports (entitlement required); native cost tracking removed; Agent Time (hours of model invocation) is measured separately for licensing; no per-chat or per-sub-agent cost display documented now.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Provider reconciliation [partial] Gateway sits in the path. Sources: coder-agents:gateway.
Compute cost [unsupported] Customer infrastructure; not shown. Sources: coder-agents:lic.
Refresh delay [partial] Spend is recorded after each model response and the AI Gateway checks current spend before each request, so brief overage is possible with concurrent requests; the refresh interval of reports and the usage indicator is not stated. Sources: coder-agents:spend.
Alerts [partial] In-app soft warning near Agent Time exhaustion. Sources: coder-agents:lic.
Export [supported] Gateway reports and API. Sources: coder-agents:spend.
Admission control queues agents beyond the concurrency limit (5 root, 10 sub-agent defaults); no CPU or RAM awareness or check deduplication.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Chats can be shared read-only with users or groups, with live streaming of updates and notifications; viewers cannot send messages, so no co-prompting. Projects are private to their owner.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Coder organizations and roles (Organization Admin, Owner, custom roles with model configuration access), multi-organization RBAC and OIDC group sync (Premium), OIDC login, SCIM 2.0 user provisioning (Premium, not fully spec-certified); agent permissions equal the submitting user; per-organization models and MCP servers.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface), Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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
project_roles: Projects are private to the owner and org-bound.
task_roles: None.
provisioning: SCIM 2.0 and IdP group sync are Premium; Coder says its SCIM implementation is not fully certified.
revocation: Remove membership; chat sharing can be disabled.
Sso [supported] OIDC authentication documented; group sync Premium. Sources: coder-agents:oidc.
Provisioning [supported] SCIM 2.0 provisioning and deprovisioning (users suspended on deactivation) and IdP group/role sync are documented as Premium features; Coder states its SCIM implementation is not fully certified. Sources: coder-agents:oidc, coder-agents:scim.
Unauthorized access [supported] RBAC applies to tools and templates. Sources: coder-agents:idx.
Revocation [partial] Remove membership; chat sharing can be disabled. Sources: coder-agents:share.
Plan scope [supported] Premium and AI Premium split in pricing. Sources: coder-agents:pricing.
Control-plane enforced model and tool configuration, plan mode, admin-approved MCP servers, template blocking for agents, per-user RBAC, network isolation templates; ACP subagent permission requests are auto-approved. The built-in agent does not prompt per command by default; control is by template, network and RBAC rather than per-call approval.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd), Coder dashboard Agents page (browser chat interface) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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: No per-command approval prompt documented for execute.
tool: Admin can configure tools and MCP servers; plan mode restricts tools.
delivery: Git operations attributed to the user.
allow: Admin approves MCP servers.
14 atomic answers
Command [unsupported] No per-command approval prompt documented for execute. Sources: coder-agents:idx.
Premium audit logging records project, instruction, skill and memory changes and user actions; chat history persisted with retention (default 30 days archived) and debug logs; export through API; chats attributed to users.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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: Plan proposals in transcript.
edits: Diff viewer.
decisions: None.
export: API endpoints.
11 atomic answers
Actions [supported] Tool calls in chat timeline. Sources: coder-agents:idx.
Approvals [partial] Plan proposals in transcript. Sources: coder-agents:idx.
In-chat diff viewer for GitHub and GitLab with PR/MR status, plan proposals for review, attachments, search by PR status; no review queue or request-changes workflow.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Keyboard [partial] Keyboard handling found only for the remote desktop (useDesktopConnection.ts remaps Cmd shortcuts); no documented review-specific shortcuts. Sources: coder-agents:mobilecode.
Narrow screen [partial] Responsive layout code for narrow viewports in AgentsPage (mobilecode); no documented narrow-screen review flow. Sources: coder-agents:mobilecode.
Retained: chat transcripts, diffs in-chat, attach_file durable attachments, file attachments with expiry, plan files, PR status; no gallery of recordings or screenshots beyond computer-use sub-agent and attachments.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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
screenshots: Computer-use sub-agent screenshots and attachments.
No built-in gate: tests and linters run through the execute tool when asked, with no required-check policy in Coder Agents; template and CI policies live outside the product.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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
tests: Agent runs tests via execute.
lint: Same.
types: Same.
ci: PR status in chat.
11 atomic answers
Tests [partial] Agent runs tests via execute. Sources: coder-agents:idx.
Git work happens in workspaces under the user identity with external auth; PRs opened by the agent are attributed to the user; chats link to a PR whose state (draft, open, merged, closed) is searchable; in-chat diff viewer; GitHub and GitLab supported.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface), Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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: Agent uses git in workspace.
worktrees: None.
comments: None.
ci_response: None.
9 atomic answers
Branches [partial] Agent uses git in workspace. Sources: coder-agents:arch.
Merging and deployment are performed through the git provider and CI under the user permissions; Coder Agents does not provide merge or deploy controls.
Scope
Surfaces: Coder dashboard Agents page (browser chat interface) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Coder core (control plane, dashboard, chatd, CLI) is AGPL-3.0 at v2.39.4; enterprise-licensed code under enterprise/ (audit, AI Gateway, aiseats) is source-available under LICENSE.enterprise. Coder Agents code lives in the AGPL tree (coderd/x/chatd); some governance features are entitlement-gated.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd), Coder dashboard Agents page (browser chat interface), Coder Agents HTTP API (/api/v2/chats) and coder CLI · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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
Premium and AI Premium features sit under the enterprise license.
Fully self-hosted: control plane, database and agent loop run on your infrastructure with no vendor service except Agent Time usage reporting (hours only) for licensed deployments; chats and data stay in your database; export through the API; standard formats not described.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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_dependencies: Agent Time hours reported to Coder for AI Premium; LLM providers.
export: API access to chats.
import: None found.
formats: Database and REST JSON.
8 atomic answers
Self hosting [supported] Entire product self-hosted; no SaaS. Sources: coder-agents:idx.
Vendor dependencies [partial] Agent Time hours reported to Coder for AI Premium; LLM providers. Sources: coder-agents:lic.
Prompts and code go from your control plane to the LLM providers you configure; chat history stays in your database; LLM keys never enter workspaces; Coder collects installation telemetry by default (opt-out with CODER_TELEMETRY_ENABLE=false) and, for licensed deployments, hourly Agent Time totals (no user data) go to Coder; retention defaults to 30 days for archived chats.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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
training: Provider policies not captured.
subprocessors: Usage hours reach Coder and Metronome; providers chosen by you.
consent_prompt: docs/admin/setup/telemetry.md states telemetry is collected by default with no consent or opt-in step; opt-out only through CODER_TELEMETRY_ENABLE=false.
15 atomic answers
Code data [supported] Control plane to configured provider. Sources: coder-agents:idx.
Prompts [supported] Same; stored in your DB. Sources: coder-agents:arch.
Credentials [supported] Provider keys in control plane only. Sources: coder-agents:idx.
Consent prompt [unsupported] docs/admin/setup/telemetry.md states telemetry is collected by default with no consent or opt-in step; opt-out only through CODER_TELEMETRY_ENABLE=false. Sources: coder-agents:telemetry.
Provider separation [supported] Provider chosen by admin. Sources: coder-agents:platform.
Community is free for self-hosting with up to five concurrently active agents; Premium is priced per user per year (amount not shown) and includes Coder Agents with the same five-agent pool; AI Premium is custom priced with unlimited agents against purchased Agent Hours. Model and infrastructure costs are separate.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Scenario costs cannot be priced publicly: Community software is free, while Premium and AI Premium prices are not published.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Maintained commercial open-source product with patch releases on 2026-11-16 across v2.39 and v2.38 lines, GitHub and Discord community support, SLA-backed support on Premium; several agents features are experimental or early access.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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: Agents core documented as available; ACP, advisor experimental; projects early access.
access_restrictions: Premium and AI Premium gate some features.
response_commitments: SLA-backed support mentioned without numbers.
Ongoing free tier: Community is free and self-hosted with up to five active Coder Agents; Premium and AI Premium paid; AI Premium trial button without stated terms.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · 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: Trial for AI Premium, terms unstated.
credits_trial: None stated.
paid_only: Not paid-only.
early_access: Projects early access; ACP experimental.
Units are stated (per user per year for Premium, Agent Hours for AI Premium) but amounts are not published; AI Premium is custom.
Scope
Surfaces: Self-hosted Coder control plane (agent loop, chatd) · Version: coder v2.39.4 (commit abe6ed41c7513a74347162adb4043f1ec6038d45, 2026-11-16); docs read at the tag 2026-11-16 · Plan: Not specified · OS: Not specified · Execution modes: Not specified · Region: Not specified · Harness: Not specified
Chat interface and API for delegating work to a self-hosted Go agent that runs inside the Coder control plane; workspaces provisioned automatically from templates; sub-agents; queueing; attachments; providers Anthropic, OpenAI, Google, Azure, Bedrock, OpenAI-compatible, OpenRouter, Vercel; no SaaS component.
Kind
documented
Locator
index.md How it works; Built-in tools; Plan mode; LLM providers; Security benefits
Admin configures an LLM provider and model per organization; needs a template; every org member can use Agents by default; CI/webhook triggering listed as API use.
In-chat diff viewer and merge request status use configured external auth for GitHub and GitLab; write_repository scope allows pushing commits and creating merge requests.
The Agents page imports isMobileViewport and useMediaQuery and applies responsive classes, so the web UI adapts to narrow viewports in code; no documentation states supported phone browsers.
Kind
source-inspected
Locator
site/src/pages/AgentsPage/hooks/useConversationEditingState.ts lines 3 and 137; useAgentChatPanelPreference.ts line 2; 91 sm:/md:/lg: Tailwind class uses across AgentsPage (rg -c)
Experimental: with chat-acp-subagents experiment, external harnesses such as Claude Code or Codex run as ACP subagents in the workspace, using the harness own credentials; tool permission requests auto-approved; not in plan mode.
Community and Premium: up to five concurrently active agents deployment-wide, queued beyond; AI Premium: unlimited concurrent agents against purchased Agent Hours; usage hours reported to Coder Tallyman/Metronome without chat data.
Kind
documented
Locator
Community and Premium licenses; AI Premium; Agent Time; usage reporting
Community free (up to 5 active Coder Agents, AI Governance); Premium priced annually per user with no amount shown (adds Coder Agents, HA, audit logging, multi-org RBAC, OIDC sync, SLA); AI Premium custom (unlimited active agents, Coder Agents API for headless agents, network-isolated agent workspaces).
Receivers must tolerate duplicate pubsub deliveries and the run loop deduplicates state updates; duplicate execution of tool calls after an interruption is not addressed.
Kind
source-inspected
Locator
coderd/x/chatd/ARCHITECTURE.md line 671 and runner.go line 120
Per-user and group monthly budgets enforced through AI Gateway (requires entitlement); native cost tracking removed; limit cap $1,000,000 per member per period.
The execute tool waits 10s by default (timeout argument up to the caller), supports run_in_background and a wait_timeout override; no approval prompt appears in the tool.
Kind
source-inspected
Locator
coderd/x/chatd/chattool/execute.go lines 34-55 and 131
SCIM 2.0 user provisioning and deprovisioning with header authentication is a Premium feature; deactivation suspends users; the docs say the implementation is not fully certified against the SCIM 2.0 specification.