dsh-claude-mcp

Hand one task to the Claude Code CLI on this machine, and get the result back cleanly — while Claude cannot touch your project files. Good for reviewing a plan, a code change or a PR, or for getting a second opinion.

What it does

• Runs one Claude Code task for you. Give it a complete, self-contained job ("review this plan", "review this change"), it runs claude -p, and brings the answer back.

• Keeps Claude out of your files. Claude reads the directories you name, but the only place it can write is its own throwaway run directory.

• Hands back a list, not a wall of text. You get Claude's final message plus a manifest of what it produced: name, size, and a sha256 checksum. Open what you need, decide for yourself, then copy the parts you approve into your project.

• One tool. Nothing else to learn and nothing else to configure.

Related MCP server: cursor-agent-sync-mcp

Requirements

Node.js 22.19 or newer, or 24 or newer, plus a working, already-signed-in claude command on the same machine.

Install

1. Append the block below to ~/.dsh/profiles/<your profile>/cordis.patch.yml, replacing every /absolute/path/... with a real absolute path:

- insert:
    - id: mcp-claude
      name: '@deepseek-ai/dsh-mcp-client'
      config:
        transport: stdio
        serverName: claude
        command: /absolute/path/to/node
        args:
          - /absolute/path/to/dsh-claude-mcp/src/server.mjs
        cwd: /absolute/path/to/your/workspace
        toolCallTimeoutMs: 900000
        failOnStartupError: true

Absolute paths matter: a desktop app launched from the Finder inherits a minimal PATH, so a bare node or a relative path will not resolve.

2. If the tool does not appear, restart the app once. A profile that reloads its patches live picks the row up on its own; the desktop profile declares no live reload.

You then have one extra tool: mcp__claude__run.

Usage

Ask me for a review, or anything else you want a second opinion on, and I will call it. What you can specify:

A run looks like this:

ok  exit=0  140.1s  model=sonnet  permission=dontAsk
artifact   <workspace>/.claude-staging/run-20261001-191307-9c44/work
final message
-------------
review written to review.md
artifacts (1) — bodies are NOT included, read them yourself
-------------
    2500 B  sha256:6676042702213315  review.md

artifact is the directory Claude was allowed to write; the run directory also holds a manifest.json recording the same facts. Open the files the list names, read them, and copy the parts you approve into your project.

If the boundary blocked something Claude wanted to do, the run says so under denied by the permission boundary; the answer may then be incomplete, so read that list first. Keep answers short and structured — say how long they should be, because very long replies get truncated.

Command line (optional)

claude-mcp run -p "Review plan/foo.md and write review.md" -m sonnet
claude-mcp runs                                 # list past runs
claude-mcp prune --older-than-days 7 --keep 5   # clean up old run directories
claude-mcp env                                  # show how Claude Code will be launched

Configuration (optional)

Safety

• Claude reads the directories you name and writes only in its own run directory. Both are enforced by Claude Code's permission rules, not by asking the model nicely; the session has no shell and no network tools at all, and your Claude Code login is used as-is. This tool never reads, stores, or forwards your credentials, and variables whose names look like credentials are not passed on.

• Nothing Claude produces enters your project by itself — you decide what to copy.

• Every run leaves a manifest, and claude-mcp prune is the only thing that deletes run directories.

Development

node --test tests/*.test.mjs

The tests are fully offline: they use a fake claude, call no model and cost nothing. The real thing lives in scripts/live-review.mjs (that one costs money).

License

MIT.