Claudexor

Name: Claudexor
Author: Anton Razzhigaev

Routes coding tasks to Claude Code, Codex, Cursor, OpenCode, Antigravity or Grok Build and tracks quota

By Anton Razzhigaev (individual) JavaScript Local · stdio Developer tools AI and ML

GitHub npm Website

53 popularity score ★ 522 ↓ 723/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 claudexor
claudexor doctor

Claude CodeTerminal

claude mcp add claudexor -- claudexor mcp serve

Claude Desktopclaude_desktop_config.json

{
  "mcpServers": {
    "claudexor": {
      "command": "claudexor",
      "args": [
        "mcp",
        "serve"
      ]
    }
  }
}

Cursor~/.cursor/mcp.json

{
  "mcpServers": {
    "claudexor": {
      "command": "claudexor",
      "args": [
        "mcp",
        "serve"
      ]
    }
  }
}

VS Code.vscode/mcp.json

{
  "servers": {
    "claudexor": {
      "type": "stdio",
      "command": "claudexor",
      "args": [
        "mcp",
        "serve"
      ]
    }
  }
}

Codex~/.codex/config.toml

[mcp_servers.claudexor]
command = "claudexor"
args = ["mcp", "serve"]

Gemini CLITerminal

gemini mcp add --scope user claudexor claudexor -- mcp serve

Settings 6

Set together: raw api
`OPENAI_API_KEY` secret	Optional: Key for the raw API route when no signed-in agent CLI is used; a stored openai secret is tried when it is unset. For raw API route.
`CLAUDEXOR_RAWAPI_KEY` secret	Optional: Takes the place of OPENAI_API_KEY for the raw API route when set. For raw API route.
`CLAUDEXOR_RAWAPI_BASE_URL`	Optional: Endpoint the raw API route calls. Default `https://api.openai.com/v1`. For raw API route.
`CLAUDEXOR_RAWAPI_MODEL`	Optional: Model the raw API route uses when none is chosen. Default `gpt-4o-mini`. For raw API route.
`CLAUDEXOR_CONFIG_DIR`	Optional: Points Claudexor at another state folder. Default `~/.claudexor/v3`.
`CLAUDEXOR_DISABLE_STORED_SECRETS`	Optional: Set to 1 to stop keys being looked up in the stored secrets file.

What it does

Claudexor sits in front of the coding agents a developer already has signed in on their own machine. It can drive Codex CLI, Claude Code, Cursor CLI, OpenCode, Antigravity and xAI's Grok Build on a grok.com subscription. A raw model API with a key is the fallback route. Its MCP server starts with `claudexor mcp serve` over stdio, so another assistant can hand work to that layer rather than to one agent.

It registers 41 tools. The main ones queue a read-only question, a plan, an agent run that edits code, a new project, or a best-of race where two or more attempts compete and a model from another vendor judges them. Each queues the job with a local daemon and returns a handle. The rest read status and results, wait on many runs at once, answer questions a run asks, send a live message to a running attempt, change its time limit, or start a follow-up. Each prompt tool also takes per-run controls, such as a web policy or a hard time limit. Threads keep a conversation with a project across turns, and upload tools bring a local file in as an attachment.

Account tools show which named subscription profiles are ready, their quota and balances, and can spend a quota reset a vendor offers. Two recovery tools inspect the daemon's journal and can set a damaged part aside.

A few things live only in the code. When a call leaves repoPath out, the project is the folder the server was started in. The raw API route defaults to the OpenAI endpoint and gpt-4o-mini. And Claudexor does not install Grok itself: it finds the grok program through CLAUDEXOR_GROK_BIN before falling back to the PATH and then ~/.grok/bin.

Since 5.0, credential-shaped prompt text and project files such as .env are passed through untouched. The changelog moves the only content check to `pr` publication, where a pinned Betterleaks scan reports findings by rule and place without echoing the matched text.

What it needs

Runtime	Node.js 24.15+
Runs on	macOS, Linux
Coding agent subscription (Claude, Codex, Cursor, Grok...)	Paid planRuns spend your own plan quota; OpenCode and Antigravity also work, and a provider API key covers the raw API route
Also	At least one vendor CLI on PATH: codex, claude, cursor-agent, opencode, agy or grok Git, for worktrees and patch delivery

Install with `npm install -g claudexor`, then check the setup with `claudexor doctor`. Sign each agent in through Claudexor rather than the bare vendor CLI. Claudexor can install some agent CLIs itself with `claudexor harness install <harness> --target local --yes --json`, pinned to an exact version under ~/.claudexor/node; Grok is the exception and has to be installed separately. On Windows only the Codex and Claude installs work. Before the first `pr` delivery, fetch the pinned publication scanner with `claudexor scanner prepare`, or with `claudexor scanner prepare --from-archive <file>` when offline. A running daemon picks it up at its next `pr`.

Limits

Limit Run tools return a handle; the answer comes later through claudexor_run_status or claudexor_run_result, since MCP Tasks are not used yet
Not covered Quota tracking covers Claude, Codex and Antigravity; Cursor's balance is only read from its dashboard
Limit If host plugin files and the CLI versions differ, mcp serve will not start until `claudexor plugin repair all` is run
Early release The ACP surface and the host plugin file layout are marked experimental; the MCP tool set is the stable contract
Early release Grok Build sign-in from the app is unverified against a real vendor session, and live checks ran only on Windows 11
Limit `pr` delivery refuses until the pinned scanner, which the npm packages leave out, is fetched with `claudexor scanner prepare`
Not covered Selected .env and key files are copied and published as-is in directory execution, and changes made in place have no full rollback
Platform The Copilot plugin runs on macOS and Linux only
Not covered Only claude, codex and cursor can host the delegated belt, one level deep with 8 sub-runs per parent by default
Not covered web=off turns off each agent's web tools but leaves network access open, and Cursor is refused under off

Run tools do not return the answer: they queue work and return a handle, and the result must be read with claudexor_run_status and claudexor_run_result. MCP Tasks are not offered. Quota tracking and account rotation cover Claude, Codex and Antigravity; Cursor has no usage source yet. If the installed host plugin files and the CLI versions differ, `mcp serve` refuses to start until `claudexor plugin repair all` is run. The ACP surface and host plugin layout are marked experimental. claudexor_thread_read fails with daemon_unavailable when no daemon is running. The portable Copilot plugin does not support Windows.

Who it suits

A fit for a developer who already pays for two or more coding agents and wants one assistant to share tasks among them, race them and keep an eye on quota, all on their own Mac or Linux box. It brings a daemon, a state folder under ~/.claudexor and Node 24, and its run tools answer through handles. Someone with a single agent and no use for races or quota tracking gains little from it.

Tools 41

23 only read, 15 change things, 3 can delete or overwrite. A ? marks an optional parameter.

claudexor_ask: Queue a read-only question about a project, with an optional deeper multi-scout passprompt: stringrepoPath?: stringharness?: stringprimaryHarness?: stringmodel?: stringeffort?: stringprocessingPreference?: stringidempotencyKey?: stringattachments?: arrayexecution?: objectdeepScan?: booleantaskGroupId?: stringweb?: stringaccess?: stringcredentialProfileId?: stringpaidBudget?: objectmaxSeconds?: integertimeLimitMode?: stringn?: integer
claudexor_plan: Queue a read-only implementation plan for a taskprompt: stringrepoPath?: stringharness?: stringprimaryHarness?: stringmodel?: stringeffort?: stringprocessingPreference?: stringidempotencyKey?: stringattachments?: arrayexecution?: objecttaskGroupId?: stringweb?: stringaccess?: stringcredentialProfileId?: stringpaidBudget?: objectmaxSeconds?: integertimeLimitMode?: stringn?: integercouncil?: boolean
claudexor_run changes: Queue an agent run that edits code; model review is off unless asked forprompt: stringrepoPath?: stringharness?: stringprimaryHarness?: stringmodel?: stringeffort?: stringidempotencyKey?: stringattachments?: arrayexecution?: objectn?: integerprotectedPathApprovals?: arraytaskGroupId?: stringweb?: stringaccess?: stringcredentialProfileId?: stringpaidBudget?: objectmaxSeconds?: integertimeLimitMode?: stringtests?: arrayreview?: boolean
claudexor_best_of changes: Queue at least two competing attempts on one task and have another vendor's model judge themprompt: stringrepoPath?: stringharness?: stringprimaryHarness?: stringmodel?: stringeffort?: stringidempotencyKey?: stringattachments?: arrayexecution?: objectn?: integertaskGroupId?: stringweb?: stringaccess?: stringcredentialProfileId?: stringpaidBudget?: objectmaxSeconds?: integertimeLimitMode?: stringtests?: array
claudexor_create changes: Queue a run that starts a new projectprompt: stringrepoPath?: stringharness?: stringprimaryHarness?: stringmodel?: stringeffort?: stringidempotencyKey?: stringattachments?: arrayexecution?: objecttaskGroupId?: stringweb?: stringaccess?: stringcredentialProfileId?: stringpaidBudget?: objectmaxSeconds?: integertimeLimitMode?: stringn?: integertests?: arrayreview?: boolean
claudexor_thread_create changes: Open a lasting thread on a project without starting a model; write turns edit the folder in place unless workspace is isolatedrepoPath: stringtitle?: stringdefaultMode?: enumworkspace?: enumworkspaceRoot?: stringcredentialProfileId?: stringprimaryHarness?: stringeligibleHarnesses?: arrayaccess?: enumidempotencyKey?: string
claudexor_thread_turn changes: Send the next turn of a thread and get a run handlethreadId: stringprompt: stringattachments?: arrayrunMode?: enumharness?: stringprimaryHarness?: stringmodel?: stringeffort?: stringcredentialProfileId?: stringaccess?: enumweb?: enummaxSeconds?: integertimeLimitMode?: enumexecution?: objectidempotencyKey?: string
claudexor_thread_read: Read a thread and its session records without starting workthreadId: string
claudexor_thread_history: Page through a thread's turns, up to 20 at a time, with answer previewsthreadId: string
claudexor_thread_prompt: Read the full stored prompt of one turn in byte-bounded partsthreadId: string
claudexor_store_identity: Report which daemon store the server is talking to, before uploads
claudexor_upload_import changes: Import one absolute local file on the server's machine as an attachment in a single keyed call
claudexor_upload_create changes: Start a multi-step upload
claudexor_upload_import_local changes: Feed a local file into an upload that is already open
claudexor_upload_status: Check how far an upload has got
claudexor_upload_finalize changes: Close an upload and get the stored resource
claudexor_upload_cancel changes: Abandon an open upload
claudexor_status: Show whether each harness is ready and which model is set
claudexor_capabilities: List what each harness and mode can do and may change
claudexor_accounts: Show credential profiles with readiness and quota; fresh probes vendors and costs morefresh?: boolean
claudexor_account_resources: Read balances, spending caps, quota and reset offers for accountsrefresh?: booleantarget?: objectmodel?: string
claudexor_account_reset deletes: Spend a quota reset an account offers, or look up an earlier reset by operation idtarget?: objectoffer_id?: stringgrant_id?: stringidempotency_key?: stringoperation_id?: string
claudexor_runs: List recent runs so a lost run id can be found
claudexor_inspect: Read one run's details and evidencerunId: string
claudexor_run_status: Read where a run stands nowrunId: string
claudexor_run_result: Read a finished run's outputrunId: string
claudexor_run_cancel deletes: Stop a run that is still goingrunId: string
claudexor_run_interactions: List the questions a run is waiting onrunId: string
claudexor_answer_interaction changes: Answer a question a run is waiting onrunId: stringinteractionId: stringanswers: array
claudexor_apply_check: Check without changes whether a run's patch would applyrunId: string
claudexor_run_wait: Wait on one run and replay its events in bounded batches (50 events, 25 s by default)
claudexor_run_event_read: Read one large run event piece by piece
claudexor_runs_wait: Watch up to 1000 run handles in one calltargets: arraywaitMs?: integer
claudexor_task_group_runs: Read one page of the runs in a task grouptaskGroupId: stringlimit?: integercursor?: stringexpectedStoreId?: string
claudexor_run_time_limit changes: Read, set, extend or remove a run's time limitrunId: stringaction: enumseconds?: integermode?: enumexpectedRevision?: integer
claudexor_run_message changes: Send a live message to the attempt a run is runningrunId: stringtext: stringidempotencyKey: string
claudexor_run_continue changes: Start a follow-up run from a finished single-candidate Agent or Ask runrunId: stringprompt?: stringidempotencyKey?: stringcontinueCarrier?: enum
claudexor_next_message changes: Draft, replace or clear the next message queued behind an Agent or Ask run
claudexor_next_message_read: Read a queued next-message draft and the settings frozen with it
claudexor_journal_recovery: Inspect, validate or export one partition of the daemon journalaction: enumpartition: string
claudexor_quarantine_journal deletes: Set a damaged journal partition aside and start a fresh one, after an exact confirmationpartition: stringexpectedFingerprint: stringconfirmation: enum

Popularity

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

★ 522 ↓ 723/wk

GitHub stars, whole repository	522	22.9 points
Stars in 30 days	+24	7 points
npm downloads, last week	723	22.9 points

It lives in the folder packages/claudexor of razzant/claudexor, which holds more than this server. Its stars are the whole repository's, so in the score they never count for more than its own downloads.

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

Where it is published

Code	razzant/claudexor/packages/claudexor ★ 522, the whole repository
npm	claudexor 5.1.0, 2026-11-08

What it can reach

Runs	On your machine
Reaches	Your code, Files on your machine, Runs shell commands, Your account at a service, The internet
Changes things	Yes: it can write, send, create or delete
Sign-in	Account credentials
Licence	MIT (npm)
Usage data	Sends none No usage data, analytics or crash reports; the telemetry folder holds local cost and latency figures only
Last code change	2026-11-08 (a release)

Everything runs locally: a daemon on a Unix socket plus a loopback HTTP API that wants a bearer token. Agent runs and thread turns edit code, in the project folder itself unless workspace=isolated is picked. Keys sit in a 0600 file and vendor tokens are never copied. Nothing is filtered on the way in. Credential-shaped prompt text reaches the agent and is stored as sent, and selected files, .env and keys included, are copied as they are. The one content check is a pinned Betterleaks scan at `pr` publication. upload_import still refuses credential-store paths such as .grok/. web=off drops each agent's web tools but does not cut network access. The daemon polls Anthropic's usage endpoint and reads the Cursor dashboard with your session cookie.

Over time

GitHub stars, whole repository 522 +24 (+5 %)

523498

9 Oct 20268 Nov 2026

npm downloads a week 723 +51 (+8 %)

1.7k567

20 Jul 20268 Nov 2026

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

Popularity score 53 −3 (−6 %)

560

9 Oct 20268 Nov 2026

What is rising across the directory

Sources

Page checked 2026-11-06.

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

[![Claudexor on EveryMCPServer](https://everymcpserver.com/badge/claudexor.svg)](https://everymcpserver.com/servers/claudexor)

Connect your client

Add in one click

Or add it by command or config

Settings 6

What it does

What it needs

Limits

Who it suits

Tools 41

Similar servers

ClaudeR

MartinLoop

Ruflo

Claw Orchestrator

Claude Code Ultimate Guide MCP

Claudish