Claw Orchestrator

Name: Claw Orchestrator
Author: Guian Fang

Runs Claude Code, Codex, Grok Build and ACP agents as sessions, with checks the agents cannot edit

By Guian Fang (individual) TypeScript Local · stdio AI and ML Developer tools

GitHub npm Website

66 popularity score ★ 602 ↓ 2.1k/wk

Connect your client

Local (npm)

Add in one click

Add to Cursor Add to VS Code Add to LM Studio

Opens the app, which asks before it adds the server.

Or add it by command or config

FirstTerminal

npm install -g @enderfga/claw-orchestrator

Claude CodeTerminal

claude mcp add claw-orchestrator -- clawo-mcp

Claude Desktopclaude_desktop_config.json

{
  "mcpServers": {
    "claw-orchestrator": {
      "command": "clawo-mcp",
      "args": []
    }
  }
}

Cursor~/.cursor/mcp.json

{
  "mcpServers": {
    "claw-orchestrator": {
      "command": "clawo-mcp",
      "args": []
    }
  }
}

VS Code.vscode/mcp.json

{
  "servers": {
    "claw-orchestrator": {
      "type": "stdio",
      "command": "clawo-mcp",
      "args": []
    }
  }
}

Codex~/.codex/config.toml

[mcp_servers.claw-orchestrator]
command = "clawo-mcp"
args = []

Gemini CLITerminal

gemini mcp add --scope user claw-orchestrator clawo-mcp

Settings 7

Set together: gateway
`ANTHROPIC_API_KEY` secret	Optional: Key for the Claude Code engine when its CLI is not signed in.
`OPENAI_API_KEY` secret	Optional: Key for the Codex engine when its CLI is not signed in.
`GEMINI_API_KEY` secret	Optional: Gemini key for the multi-model proxy route, which clawo-mcp does not serve. For multi-model proxy.
`CLAWO_MCP_TOOLS`	Optional: Comma-separated tool names to advertise. Unset exposes all 79.
`CLAWO_NO_EMBEDDED_SERVER`	Optional: Keeps the HTTP control plane on port 18796 closed. clawo-mcp sets it to 1 when unset. Default `1`.
`GATEWAY_URL`	Optional: Address of an OpenClaw or Anthropic-style gateway for the proxy route, used with GATEWAY_KEY. For multi-model proxy.
`GATEWAY_KEY` secret	Optional: Key for the gateway in GATEWAY_URL. For multi-model proxy.

What it does

Claw Orchestrator is a local runtime that keeps coding-agent command-line tools running as long-lived sessions and puts them behind one tool surface. Out of the box it drives Claude Code, Codex, Antigravity (agy), Grok Build and OpenCode. It can also run any agent that speaks the Agent Client Protocol over stdio as `engine: 'acp'`, with bundled presets for Kimi Code and dsh, and any other coding CLI can be wired up as a custom engine. The MCP server, `clawo-mcp`, ships in the npm package next to the `clawo` CLI and the `clawo-acp` agent, and it registers 79 tools.

From a host you can start a session on one engine, keep sending it work, and hand the conversation to another engine halfway through. You can also put one question to several engines at once, or set a council of agents loose in separate git worktrees until they agree. Workflows are checkpointed, so a run picks up again after a restart. If you attach an acceptance contract (test commands, HTTP probes, screenshots, diff rules, file checks), the runtime runs those checks itself and marks the run verified, refuted or unverified from the results instead of from what the agent claims. The `goal` workflow template (also `clawo goal`) puts a Codex /goal on one codex-app thread and then verifies and repairs on that same thread. Each turn lands in a ledger with tokens and cost, and `maxBudgetUsd` stops further sends once a session has spent its cap.

A paused autoloop run can be picked up again with `autoloop_resume`.

Some defaults show up only in the tool schemas in src/index.ts. Council agents get `bypassPermissions` unless you pass `defaultPermissionMode`, and a council runs up to 15 rounds with 50 tool turns per agent per round by default. A fan-out takes at most 16 agents. `project_purge` only reports what it would delete until you pass `dry_run=false`. On launch `clawo-mcp` sets `CLAWO_NO_EMBEDDED_SERVER=1` if it is not already set, so the HTTP dashboard on port 18796 that `clawo serve` opens stays closed.

What it needs

Runtime	Node.js 22+
Runs on	macOS, Windows, Linux
A coding engine vendor (Anthropic, OpenAI, xAI or other)	Paid planAt least one engine CLI installed and signed in. Each engine bills on its own vendor account.
Also	At least one engine CLI on PATH: claude, codex, agy, grok, opencode, or an ACP agent such as kimi or dsh

Install the package globally (`npm install -g @enderfga/claw-orchestrator`) so `clawo-mcp` is on PATH. The engine CLIs are not included and must be installed and signed in separately. MCP hosts do not pass on your shell environment, so put any API key or PATH an engine needs in the host's env block. Per the MCP guide, `CLAUDE_BIN`, `CODEX_BIN` and similar point at an engine binary that is not on PATH.

Limits

Not covered Each MCP host starts its own clawo-mcp process, so sessions started from one host are not visible from another.
Limit If none of the names in CLAWO_MCP_TOOLS match a tool, the server advertises no tools and only prints a warning to stderr.
Not covered Autoloop roles cannot use the acp engine, and ultrareview reviewers cannot use acp or custom engines (per src/index.ts).
Not covered Fan-out, council and workflow agents take an ACP agent by preset id only. An inline ACP config is accepted from a local caller only.
Not covered The goal workflow template and the codex_goal_* tools work on codex-app sessions only.
Cost Every engine bills its own vendor account. maxBudgetUsd relies on estimated cost for engines that estimate token counts.
Platform A read-only sandboxMode on Grok is refused on systems without Seatbelt or Landlock. The acp engine refuses read-only outright.
Limit A custom engine must map permission modes in its config, or its session refuses to start.
Not covered The dashboard and the OpenAI-compatible endpoint need clawo serve. The MCP server does not open them.

With 79 tools, a host with a small context gets crowded. `CLAWO_MCP_TOOLS` limits the list, but if none of its names match, the server advertises no tools at all and prints a warning to stderr. Each MCP host starts its own `clawo-mcp` process, so sessions started from one host (or one Codex session) are not visible from another. `maxBudgetUsd` relies on cost estimates for engines that estimate token counts. A read-only `sandboxMode` on Grok is refused on platforms without Seatbelt or Landlock, and a custom engine must map permission modes or its session will not start. The dashboard and the OpenAI-compatible endpoint need `clawo serve`, not the MCP server.

Who it suits

It suits developers who already use two or more coding CLIs and want one assistant to spread work across them, compare answers, or accept a change only after real checks pass. The agents it starts edit code and run commands on your machine, and councils skip permission prompts by default, so point it at a project folder or worktree you can afford to have changed. If you only use one vendor's agent and do not need cross-checks, that agent's own subagents are simpler and spare your host a 79-tool list. If you already run OpenClaw, the project's docs suggest installing it as an OpenClaw plugin instead of through MCP.

Tools 60 of 79

21 only read, 36 change things, 3 can delete or overwrite. A ? marks an optional parameter.

session_start changes: Start a long-lived coding-agent session on a chosen engine, model and folder, including ACP agentsname?: stringcwd?: stringengine?: enummodel?: stringpermissionMode?: enumeffort?: enumallowedTools?: arraydisallowedTools?: arrayaddDir?: arraymaxTurns?: numbermaxBudgetUsd?: numbersystemPrompt?: stringappendSystemPrompt?: stringworktree?: anymcpConfig?: stringultracode?: booleansandboxMode?: enumcustomEngine?: anyacpAgent?: anyresumeSessionId?: string
session_send changes: Send a message into a running session and wait for its replyname: stringmessage: stringeffort?: enumplan?: booleantimeout?: numberstream?: boolean
session_handoff changes: Replay a live conversation onto another engine or model in the same foldername: stringengine: enummodel?: stringnewName?: stringmessage?: stringmaxChars?: numbercustomEngine?: anyacpAgent?: any
session_stop changes: Stop a running sessionname: string
session_list: List the sessions this process holds
sessions_overview: Summarise every session at once
coding_session_status: Report the state of one sessionname: string
session_grep: Search a session's history for a patternname: stringpattern: stringlimit?: number
session_compact changes: Compact a session's context, optionally with your own summaryname: stringsummary?: string
session_update_tools changes: Change which tools a session's agent may usename: stringallowedTools?: arraydisallowedTools?: arrayremoveTools?: arraymerge?: boolean
session_switch_model changes: Move a session to another modelname: stringmodel: string
project_purge deletes: Delete Claude Code's stored state for a project; a dry run unless dry_run is falsepath?: stringall?: booleandry_run?: boolean
claude_goal_set changes: Set a goal on a Claude sessionname: stringobjective: stringtimeout?: number
claude_goal_clear changes: Clear the goal on a Claude sessionname: stringtimeout?: number
claude_goal_status: Read the goal status of a Claude sessionname: stringtimeout?: number
claude_agents_list: List the Claude Code subagents defined for a folder or everywhereall?: booleancwd?: string
codex_resume changes: Resume an earlier Codex conversation with a new messagesession_id?: stringlast?: booleanmessage: stringcwd?: stringmodel?: stringtimeout?: number
codex_review: Run a Codex code review of uncommitted changes, a branch or a commitprompt?: stringcwd?: stringuncommitted?: booleanbase?: stringcommit?: stringtitle?: stringmodel?: stringtimeout?: number
codex_goal_set changes: Set a goal on a codex-app session, with an optional token budgetname: stringobjective: stringtokenBudget?: number
codex_goal_get: Read the goal of a codex-app session without sending a turnname: string
codex_goal_pause changes: Pause goal work on a codex-app sessionname: string
codex_goal_resume changes: Resume a paused goal on a codex-app sessionname: string
codex_goal_clear changes: Remove the goal from a codex-app sessionname: string
codex_interrupt changes: Cancel the turn a codex-app session is runningname: string
codex_steer changes: Add input to a codex-app turn that is still runningname: stringmessage: string
codex_fork changes: Branch a codex-app thread into a new threadname: string
codex_rollback deletes: Drop the last turns from a codex-app threadname: stringnumTurns: number
codex_models: List the models a codex-app session can usename: string
codex_thread_list: List Codex threads, with search and folder filtersname: stringsearchTerm?: stringcwd?: stringarchived?: booleancursor?: stringlimit?: number
fanout_start changes: Send one task to up to 16 engine or model agents in paralleltask: stringprojectDir: stringagents: arraysynthesize?: booleansynthesisModel?: stringsynthesisEngine?: enumsynthesisAcpAgent?: stringagentTimeoutMs?: numbermaxTurnsPerAgent?: numbermaxBudgetUsd?: number
fanout_status: Collect a fan-out's answers and progressid: string
fanout_abort changes: Cancel a fan-outid: string
coding_agents_list: List the engine CLIs found on this machinecwd?: string
team_list: List the members of a session's agent teamname: string
team_send changes: Send a message to one teammate in a session's agent teamname: stringteammate: stringmessage: string
council_start changes: Start agents in separate git worktrees that work in rounds until they agreetask: stringprojectDir: stringagents?: arraymaxRounds?: numberagentTimeoutMs?: numbermaxTurnsPerAgent?: numbermaxBudgetUsd?: numberdefaultPermissionMode?: enum
council_status: Read a council's progressid: string
council_abort changes: Stop a running councilid: string
council_inject changes: Add a message into a running councilid: stringmessage: string
council_review: Show a finished council's result for reviewid: string
council_accept deletes: Accept a council's work and remove its worktrees, branches and plan filesid: string
council_reject changes: Reject a council's work and write the feedback into its planid: stringfeedback: string
session_send_to changes: Send a message from one session to anotherfrom: stringto: stringmessage: stringsummary?: string
session_inbox: Read the messages waiting for a sessionname: stringunreadOnly?: boolean
session_deliver_inbox changes: Deliver a session's queued messages into itname: string
ultraplan_start changes: Start a planning run in a foldertask: stringcwd?: stringmodel?: stringtimeout?: number
ultraplan_status: Read a planning run's stateid: string
ultrareview_start changes: Start a code review by several reviewer agentscwd: stringagentCount?: numbermaxDurationMinutes?: numbermodel?: stringfocus?: stringengines?: array
ultrareview_status: Read a review run's stateid: string
autoloop_start changes: Start a planner, coder and reviewer loop on a workspacerun_id: stringworkspace: stringplanner_engine?: enumplanner_model?: stringplanner_effort?: enumplanner_custom_engine?: anycoder_engine?: enumcoder_model?: stringcoder_effort?: enumcoder_custom_engine?: anyreviewer_engine?: enumreviewer_model?: stringreviewer_effort?: enumreviewer_custom_engine?: anysend_timeout_ms?: numberactivity_lease_ms?: numberautoloop_hard_timeout_ms?: number
autoloop_chat changes: Talk to an autoloop's plannerrun_id: stringtext: string
autoloop_status: Read an autoloop's staterun_id: string
autoloop_list: List autoloops
autoloop_reset_agent changes: Restart one agent inside an autolooprun_id: stringagent: enumforce?: booleaneager_restart?: boolean
autoloop_stop changes: Stop an autolooprun_id: stringreason?: string
autoloop_resume changes: Resume a paused autoloop run, optionally raising its send timeoutrun_id: stringsend_timeout_ms?: numberpending_dispatch_id?: stringplannerCustomEngineRef?: stringcoderCustomEngineRef?: stringreviewerCustomEngineRef?: string
workflow_start changes: Start a durable workflow from a spec or a template (solve, council, fanout, goal), with an optional acceptance contractspec?: objecttemplate?: enumtask?: stringagents?: arrayreviewers?: arrayimplementer?: objecthumanGate?: booleanmaxRounds?: numbermaxRepairs?: numbermodel?: stringtokenBudget?: numberlauncher?: arraycwd?: stringrunId?: stringcontract?: object
workflow_status: Read a workflow run's state and recent eventsrunId: stringevents?: number
workflow_resume changes: Resume a workflow run from its last checkpointrunId: stringagentCustomEngineRefs?: object
verify_run changes: Run an acceptance contract's checks in a folder and return the evidencecwd: stringbaseSha?: stringlabel?: stringcontract: object

Popularity

66 popularity score, on our list (the bar is 36)

★ 602 ↓ 2.1k/wk

GitHub stars	602	33.4 points
Stars in 30 days	+15	6 points
npm downloads, last week	2,137	26.6 points

Measured 2026-11-08. How the score works.

Where it is published

Code	enderfga/claw-orchestrator ★ 602
npm	@enderfga/claw-orchestrator 7.20.0, 2026-11-04

What it can reach

Runs	On your machine
Reaches	Your code, Runs shell commands, Files on your machine, The internet
Changes things	Yes: it can write, send, create or delete
Sign-in	Account credentials
Licence	MIT (npm)
Last code change	2026-11-04

The server runs on your machine and starts coding-agent CLIs, including any ACP agent you name, as subprocesses in the folders you give it. Those agents edit files and run commands. Acceptance contracts run shell commands and HTTP probes, and councils create and delete git worktrees and `council/*` branches. Council agents run with `bypassPermissions` unless `defaultPermissionMode` says otherwise. `ultraapp_*` tools build web apps and start or stop their containers, and `project_purge` and `ultraapp_delete` remove stored state. Prompts and code go to each engine's vendor under that vendor's account. The server needs no key of its own and relies on the engine CLIs' sign-ins or keys passed in env. Tools carry MCP read-only and destructive hints from src/mcp-annotations.ts.

Over time

GitHub stars 602 +15 (+3 %)

605587

9 Oct 20268 Nov 2026

npm downloads a week 2.1k +230 (+12 %)

3.5k90

11 May 20268 Nov 2026

Before the site measured it, read from npm's daily download counts.

Popularity score 66 +9 (+16 %)

6655

9 Oct 20268 Nov 2026

What is rising across the directory

Sources

Page checked 2026-11-04.

Badge for your README

Shows this server's rank in its category, and stars for a server that runs on your machine. It updates every hour.

MarkdownREADME.md

[![Claw Orchestrator on EveryMCPServer](https://everymcpserver.com/badge/claw-orchestrator.svg)](https://everymcpserver.com/servers/claw-orchestrator)

Connect your client

Add in one click

Or add it by command or config

Settings 7

What it does

What it needs

Limits

Who it suits

Tools 60 of 79

Similar servers

Claudexor

Clawdi

Claudish

autocontext

AgentRQ

Crewly