Clanker
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Clanker run /clanker:codex to fix the failing tests"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Clanker
Claude Code plugin + stdio MCP server for running external coding agents as Claude-owned background tasks.
Clanker exposes these slash commands:
Command | Backend |
| Codex ACP |
| Grok ACP |
| Opencode with GLM |
| Opencode with DeepSeek |
| Opencode with Kimi |
| Opencode free model |
| Advanced Opencode model override |
Default mode is background: the Claude Agent task owns the visible bottom task row, while the MCP server owns the ACP backend process. Use --wait to run foreground.
Layout
.
├── .claude-plugin/marketplace.json
├── plugin/
│ ├── .claude-plugin/plugin.json
│ ├── .mcp.json
│ ├── agents/
│ ├── commands/
│ └── dist/clanker-mcp.mjs
├── src/
├── test/
├── scripts/
└── package.jsonRelated MCP server: openai-agents-mcp
Install
claude plugin marketplace add /path/to/ClankerHouse
claude plugin install clanker@clankerThen restart Claude Code or run /reload-plugins.
The active Claude config should point here:
"enabledPlugins": {
"clanker@clanker": true
},
"extraKnownMarketplaces": {
"clanker": {
"source": {
"source": "directory",
"path": "/path/to/ClankerHouse"
}
}
}Development
npm ci
npm run bundle
npm test
npm run typechecknpm run bundle writes plugin/dist/clanker-mcp.mjs, the self-contained server bundle Claude loads from its plugin cache.
Runtime Config
Env | Default | Meaning |
|
| Silence before a running turn is flagged as suspected stalled. |
|
| Hard per-turn ceiling before the subprocess is killed and the turn becomes |
|
| ACP initialize + session/new timeout. |
|
| Idle session TTL before reaping. |
|
|
|
| unset |
|
| server cwd | Base repo for managed worktrees. |
|
| Artifact and managed worktree roots. |
Available Tools
7 toolsclanker_cancelCancel a Clanker's in-flight turnADestructiveIdempotent
Send ACP session/cancel to the Clanker backend. Returns {id, status}. No-op if the session is idle.
Inputs: id (required).
clanker_dispatchDispatch to a Clanker and block until the turn completesADestructive
Convenience path = clanker_dispatch_start + loop clanker_wait until the turn is terminal. Returns the terminal WaitResult {status, final_message, touched_files, plan_final}. For long tasks prefer clanker_dispatch_start + clanker_wait so the caller controls polling and avoids MCP request timeouts.
Inputs: lane (required), prompt (required), cwd, model, +3 more.
clanker_dispatch_startStart a Clanker turn (non-blocking)ADestructive
Spawn/handshake a Clanker and start a prompt turn, returning {id} immediately. Poll progress with clanker_wait(id). Setup errors (unknown backend, worktree creation) fail here; runtime errors surface via clanker_wait.
Inputs: lane (required), prompt (required), cwd, model, +3 more.
clanker_listList active Clanker sessionsARead-onlyIdempotent
Overview of live Clankers: [{id, lane, state (working|idle|stalled), idle_ms, turns_count, plan_summary, suspected_stall}]. Reaped/closed sessions are omitted.
Inputs: no named parameters.
clanker_promptContinue an existing Clanker session with a new turnADestructive
Start a new prompt turn on an already-open session (persistent-session reuse). Returns {id}; poll with clanker_wait(id). Errors if the session was reaped/closed or a turn is already running.
Inputs: id (required), prompt (required).
clanker_statusCheap status of a Clanker runARead-onlyIdempotent
Return {status, plan (checkbox counts + current step), tool_calls, last_event_age_ms, suspected_stall}. Does not wait. suspected_stall flags a running turn silent past the stall threshold.
Inputs: id (required).
clanker_waitLong-poll a Clanker runARead-only
Wait up to timeout_ms (default 30000, cap 55000) for new events or completion. Returns {status, digest, plan_summary, last_event_age_ms, suspected_stall}; when status is terminal also {final_message, touched_files, plan_final}. digest is a human-readable summary of events since the previous wait — tool titles, file writes, plan check changes, key message sentences. Quiet mode (default on): only wakes before the deadline on a plan/status change, a tool error, a suspected stall, or a terminal state — trivial chatter (a tool_call starting, a file-location echo, a message-chunk fragment) does not cut the wait short, so callers no longer need to repoll tightly just because the run is reading/grepping. Pass quiet:false for the old any-event wake-up.
Inputs: id (required), quiet, timeout_ms.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 7 tools
Each tool has a clearly distinct purpose: cancel, dispatch (convenience wrapper), dispatch_start (initiate), list, prompt (existing session), status, wait. No overlap despite some convenience aliasing.
All tool names follow the consistent 'clanker_verb' pattern in snake_case. Verbs are descriptive and predictably relate to actions (cancel, dispatch, list, prompt, status, wait).
7 tools cover the full session management lifecycle without being excessive. Each tool addresses a necessary operation, from starting to polling to cancellation.
Core CRUD-like operations are covered: start a turn, wait, check status, list sessions, cancel. Missing an explicit 'delete or close session' tool, but cancel may suffice. Slight gap for cleanup.
Maintenance
Related MCP Connectors
Shared memory MCP server for Claude Code, Codex, Gemini CLI and Cursor. Cited, editable.
Real-time chat for AI agents. Claude Code, Cursor, Cline and Codex join channels over MCP.
Real-time chat hub for AI agents — Claude Code, Cursor, Cline, Codex over MCP or REST.
Augments MCP Server - A comprehensive framework documentation provider for Claude Code
Related MCP Servers
- AlicenseAqualityBmaintenanceMCP server that delegates Claude Code subagents to alternative backends such as local models, DeepSeek, or AWS Bedrock, while keeping your Claude Code orchestrator session intact.519MIT
- AlicenseAqualityDmaintenanceMCP server that bridges OpenAI's Agents SDK with Claude Code, enabling web search, file search, and computer use capabilities directly in your development environment.29 npm1MIT
- AlicenseNot gradedqualityDmaintenanceMCP server that enables Claude Code to communicate with other Claude Code agents over HTTP, allowing users to ask questions about remote codebases or delegate coding tasks.MIT
- AlicenseNot gradedqualityDmaintenanceAn MCP server that bridges Claude Desktop with Claude Code, allowing users to delegate tasks to Claude Code directly from Claude Desktop conversations, supporting both synchronous and background execution with session reuse.8 npmMIT