Claudish

Name: Claudish
Author: MadAppGang

MCP mode of the Claudish CLI: send prompts to non-Anthropic models and run them as Claude Code sessions

By MadAppGang TypeScript Local · stdio AI and ML Developer tools

GitHub npm Website

58 popularity score ★ 1.0k ↓ 1.9k/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. Then replace each <NAME> in its settings with your value.

Or add it by command or config

It needs OPENROUTER_API_KEY. Replace each <NAME> with your value.

FirstTerminal

npm install -g claudish

Claude CodeTerminal

claude mcp add claudish -e OPENROUTER_API_KEY=<OPENROUTER_API_KEY> -- claudish --mcp

Claude Desktopclaude_desktop_config.json

{
  "mcpServers": {
    "claudish": {
      "command": "claudish",
      "args": [
        "--mcp"
      ],
      "env": {
        "OPENROUTER_API_KEY": "<OPENROUTER_API_KEY>"
      }
    }
  }
}

Cursor~/.cursor/mcp.json

{
  "mcpServers": {
    "claudish": {
      "command": "claudish",
      "args": [
        "--mcp"
      ],
      "env": {
        "OPENROUTER_API_KEY": "<OPENROUTER_API_KEY>"
      }
    }
  }
}

VS Code.vscode/mcp.json

{
  "servers": {
    "claudish": {
      "type": "stdio",
      "command": "claudish",
      "args": [
        "--mcp"
      ],
      "env": {
        "OPENROUTER_API_KEY": "<OPENROUTER_API_KEY>"
      }
    }
  }
}

Codex~/.codex/config.toml

[mcp_servers.claudish]
command = "claudish"
args = ["--mcp"]

[mcp_servers.claudish.env]
OPENROUTER_API_KEY = "<OPENROUTER_API_KEY>"

Gemini CLITerminal

gemini mcp add --scope user -e OPENROUTER_API_KEY=<OPENROUTER_API_KEY> claudish claudish -- --mcp

Settings 10

Set together: meta
`OPENROUTER_API_KEY` secret	Required: OpenRouter key from the documented setup. Another provider key such as GEMINI_API_KEY, OPENAI_API_KEY or META_API_KEY, a `claudish login` subscription or a local model server can be used instead.
`CLAUDISH_MCP_TOOLS`	Optional: Tool group to expose: all, low-level, agentic, session or debug. Any other value exposes all tools. Default `all`.
`GEMINI_API_KEY` secret	Optional: Google Gemini key for g@ and google@ models.
`OPENAI_API_KEY` secret	Optional: OpenAI key for oai@ models.
`MOONSHOT_API_KEY` secret	Optional: Kimi (Moonshot) key for kimi@ models. KIMI_API_KEY also works.
`ANTHROPIC_API_KEY` secret	Optional: Anthropic key handed to the internal proxy.
`CLAUDISH_DEBUG`	Optional: 1 or true writes a debug log to logs/ in the working directory.
`CLAUDISH_TELEMETRY`	Optional: 1 sends error reports without asking. 0, false or off turns reporting off. For report_error.
`META_API_KEY` secret	Optional: Meta Model API key for metered meta@ models. The muse@ subscription signs in with `claudish login muse` and never uses this key.
`META_BASE_URL`	Optional: Endpoint for meta@ models. Default `https://api.meta.ai`.

What it does

Claudish is first of all a command-line wrapper that runs Claude Code on other models through a local Anthropic-compatible proxy. It can route to OpenRouter, Gemini, OpenAI, Kimi, GLM, MiniMax, Meta, Vertex AI and many more, or to a local Ollama, LM Studio, vLLM or MLX server. Start it with `claudish --mcp` and the same program speaks MCP over stdio, so an assistant like Claude Code can pass a question or a whole task to another model without leaving the conversation.

At v14.7.0 the server registers 16 tools in four groups. The low-level group sends one prompt to one model and lists or searches models. The agentic group runs several models on one task with anonymised answers and optional blind judging (`team`), shows which provider would serve a model (`preflight`), reports the remaining allowance of a local Codex account pool, and files error reports. The session group starts full Claude Code sessions on an external model in headless terminal panes, then lets you type into them, read their screen and output, resize, list, diagnose and cancel them. The debug group holds `compare_models`, which sends one prompt to several models in turn.

`CLAUDISH_MCP_TOOLS` picks the group: `all` (the default), `low-level`, `agentic`, `session` or `debug`. The old value `channel` was renamed to `session` and now selects every tool with a warning on stderr. Progress reaches the client as ordinary progress notifications, and the server no longer announces a channel capability. Inside a pane that claudish itself started, the server drops to the four one-shot tools, so sessions cannot start more sessions.

What it needs

Runtime	Node.js 18+
Runs on	macOS, Windows, Linux
OpenRouter	Paid planKey used in the documented setup. Another provider key, a `claudish login` subscription or a local model server can stand in for it.
Also	Claude Code CLI on PATH for team, create_session and Claude model names magmux 0.18.0+ for team and the session tools (bundled on macOS and Linux)

Node.js 18+ or Bun 1.0+, per the package's engines field. Install with `npm install -g claudish`. Provider keys go in the client's env block. A `.env` in the working directory is read at startup but only for provider keys and model settings (`CLAUDISH_MODEL`, `CLAUDISH_MODEL_*`), so put other settings in the shell or the env block. Subscriptions sign in with `claudish login` (codex, antigravity, grok, kimi, or muse for Meta's Muse Code). The pane tools need magmux 0.18.0 or newer, which ships with the npm package on macOS and Linux and otherwise comes from `brew install MadAppGang/tap/magmux`.

Limits

Platform `team` and the session tools do not run on Windows and fail with `magmux_unavailable` when magmux is missing. The other tools work on every platform.
Size cap At most 48 live panes per user across all claudish processes. Past that, a new run or session fails with `pane_limit`.
Not covered `search_models` searches only OpenRouter's listing, so a model missing from it may still be routable through another provider.
Not covered In a session claudish started, only `run_prompt`, `compare_models`, `list_models` and `search_models` are offered, whatever CLAUDISH_MCP_TOOLS says.
Not covered `allowance_status` only covers a Codex account pool. Without one it returns null with a reason.
Cost Calls to cloud models bill the provider account behind the key or draw on a subscription quota.
Size cap A session's events.jsonl record is capped at 4 MB.

`team`, `create_session` and the other pane-based features do not work on Windows and refuse with `magmux_unavailable` when magmux is missing; the other tools work on every platform. At most 48 live panes per user across all claudish processes. `team` session paths must sit inside the current working directory. `search_models` only covers OpenRouter's listing, so a model missing from it may still be routable. Session record files are capped (events.jsonl at 4 MB). The MCP guide's configuration example uses a `claudish-mcp` command, but the npm package (v10.6.1) installs only `claudish`; use `claudish --mcp` or `npx claudish@latest --mcp`.

Who it suits

Suits developers already in Claude Code who want a second opinion from GPT, Gemini, Grok, Kimi, Muse or a local model mid-session, or who want several models to try one task and be judged blind by others. The team and session tools need macOS or Linux plus the Claude Code CLI, so on Windows it is limited to single prompts and model lookup. Claudish is made by MadAppGang, not by Anthropic or any of the providers it calls. Every cloud call costs money or quota, so anyone who wants to spend nothing can point it at Ollama or LM Studio instead.

Tools 16

10 only read, 6 change things. A ? marks an optional parameter.

run_prompt: Sends one prompt to one model and returns the answer, with no tools or session. A Claude model name runs once on your own Claude Code login.model: stringprompt: stringsystem_prompt?: stringeffort?: enumuser_required_provider?: booleanuser_required_model?: boolean
list_models: Lists the models claudish recommends for coding and the suggested reasoning effort per model and task type.
search_models: Searches OpenRouter's model listing by name, provider or capability and matches results to claudish's own catalog.query: stringlimit?: number
compare_models: Debug tool that sends the same prompt to several models one after another and shows the answers side by side.models: arrayprompt: stringsystem_prompt?: string
preflight: For a list of models, says which provider would serve each, whether it is metered or subscription, and whether it answers now.models: arrayprobe?: booleantimeout_ms?: number
team changes: Runs several models on one task as Claude Code panes with anonymised outputs and optional blind judging, plus status, list, capture, resize and cancel modes.mode: enumpath?: stringmodels?: arrayjudges?: arrayefforts?: objectinput?: stringinput_file?: stringrun_id?: stringslot?: stringsince_seq?: numberspans?: booleancols?: integerrows?: integeruser_required_provider?: arrayuser_required_model?: arrayrequire_pattern?: stringmin_output_bytes?: numberagent?: stringclaude_flags?: string
report_error changes: Posts a sanitised claudish error report, optionally with local log and session files, to the developers. Meant to be called only after the user agrees.error_type: enummodel?: stringcommand?: stringstderr_snippet?: stringexit_code?: numbererror_log_path?: stringsession_path?: stringadditional_context?: stringauto_send?: boolean
allowance_status: Reports the local Codex account pool's remaining allowance, burn rate and projected gaps, without contacting any vendor.provider?: enumhorizon_hours?: numbersections?: array
create_session changes: Starts a Claude Code session on an external model in a headless pane, one-shot with a prompt or interactive without one.model: stringprompt?: stringwork_dir?: stringsetup_from?: stringtimeout_seconds?: integeragent?: stringclaude_flags?: stringeffort?: enumuser_required_provider?: booleanuser_required_model?: boolean
send_input changes: Queues a prompt for a running session, or delivers it at once when the session is waiting.session_id: stringtext: string
get_output: Returns the assistant text from each finished turn of a session.session_id: stringtail_lines?: number
cancel_session changes: Stops a session. A second call changes nothing.session_id: string
list_sessions: Lists sessions with their state, activity and idle time, finished ones on request.include_completed?: boolean
get_diagnostics: Explains a session after the fact: final screen, upstream errors, model chain, accounting and record paths.session_id: stringevent_limit?: number
capture_session: Reads a session's current terminal screen, with a sequence number for cheap polling.session_id: stringsince_seq?: numberspans?: boolean
resize_session changes: Changes a session's terminal size, clamped to at least 40 columns by 24 rows.session_id: stringcols: integerrows: integer

Popularity

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

★ 1.0k ↓ 1.9k/wk

GitHub stars, whole repository	1,014	26.2 points
Stars in 30 days	+9	5 points
npm downloads, last week	1,905	26.2 points

It lives in the folder packages/cli of madappgang/claudish, 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	madappgang/claudish/packages/cli ★ 1.0k, the whole repository
npm	claudish 14.16.0, 2026-11-07

What it can reach

Runs	On your machine
Reaches	The internet, Your account at a service, Your code, Files on your machine, Runs shell commands
Changes things	Yes: it can write, send, create or delete
Sign-in	An API key or token
Licence	MIT (npm)
Usage data	Only if you turn it on (`CLAUDISH_TELEMETRY`) Sanitised error reports: version, error type, model, command, stderr, platform and selected session files, with keys, paths and emails stripped.
Last code change	2026-11-07

Runs on your machine as a stdio process and starts a local proxy on a free port between 10000 and 19999 on first use. Prompts, and any code or text in them, go to the cloud providers you configure, billed to your keys or drawn from your subscription logins. `create_session` and `team` launch full Claude Code sessions on the external model in a working directory you choose (the current one by default). What those sessions may edit or run is set by Claude Code's permissions, the `claude_flags` you pass, and the .claude settings inherited through `setup_from`. Session transcripts, screens and event logs are written to disk. `report_error` can read log and session files you point it at and posts a sanitised report to a Google Cloud Functions endpoint run by the developers.

Over time

GitHub stars, whole repository 1.0k +9 (+1 %)

1.0k1.0k

13 Oct 20268 Nov 2026

npm downloads a week 1.9k +1.6k (+573 %)

5.9k283

22 Dec 20258 Nov 2026

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

Popularity score 58 −2 (−4 %)

6045

13 Oct 20268 Nov 2026

What is rising across the directory

Sources

Page checked 2026-10-29.

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

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

Connect your client

Add in one click

Or add it by command or config

Settings 10

What it does

What it needs

Limits

Who it suits

Tools 16

Similar servers

Crewly

Claudexor

Codeman

LLM Checker

ReevesAgents

Claw Orchestrator