codeJung-mcp

An MCP server that exposes the self-hosted codeJung code-review service as tools any MCP client can call — Claude Code, Claude Desktop, Cursor, Codex CLI, Windsurf, and others.

It is a thin client: it depends only on codeJung's stable /v1 HTTP API, not on any codeJung internals. The service binds to loopback on its host, so every call is executed on that host over SSH — the service token is read there and never leaves the machine.

Target API

Speaks codeJung's /v1 job API (POST /v1/jobs, GET /v1/jobs/{id}, GET /v1/jobs/{id}/result). If codeJung ever breaks /v1, bump this client to match. The quickest health check is the smoke test in Verifying.

Related MCP server: tai-mcp-github

Tools

Blocking vs non-blocking

review_pr / review_dir wait at most wait_secs for the review, emitting progress to stderr, then return {"status":"running","jobId":...} if it hasn't finished — they never block indefinitely. Fetch a still-running result later with get_review(job_id). Set wait_secs=0 to return as soon as the job is submitted. For clients with short tool-call timeouts, prefer the non-blocking pattern: submit_review → get_review.

Requirements

• python3 ≥ 3.10, pip install -r requirements.txt (installs mcp)

• openssh client + passwordless SSH to the codeJung host

• rsync on both machines (only needed for review_dir)

• codeJung service running on the host (~/codeJung/deploy, api + worker)

Configuration (env vars)

review_dir requires the worker container to bind-mount the staging dir. In deploy/docker-compose.yml the worker must have: - ${CODEJUNG_REVIEW_STAGING:-/home/you/cj-review-staging}:/review-staging:ro

Install per client

The same stdio server works with any local MCP client. Ready-made config snippets are in client-configs/.

Claude Code

claude mcp add codejung -s user -- python3 /path/to/codeJung-mcp/codejung_mcp.py

Cursor / Claude Desktop / Windsurf

Merge the codejung entry from client-configs/mcp.json into the client's MCP config (paths listed in client-configs/README.md). Edit the absolute path to match where you cloned this repo. MCP tools run in Agent mode, and the selected model is irrelevant — the client (not the model) drives the server, so it works with any tool-calling model (GPT, Claude, Gemini, Grok, …).

Codex CLI

Merge client-configs/codex-config.toml into ~/.codex/config.toml.

Verifying

pip install -r requirements.txt
# smoke test — should print the tool names:
python3 -c "import asyncio, codejung_mcp as m; \
print([t.name for t in asyncio.run(m.mcp.list_tools())])"
# end-to-end — the host's models must be reachable:
python3 -c "import codejung_mcp as m; print(m.review_pr('https://github.com/OWNER/REPO/pull/N')['status'])"

Usage in a session

"review https://github.com/owner/repo/pull/123 with codejung" "use codejung to review the directory ./my-service"

Security notes

• The service token is read on the host and never transits to the client.

• PR URLs and job IDs are strictly validated before being interpolated into the remote command, to prevent shell injection.

• review_dir copies the target directory to the host (excluding .git, node_modules, virtualenvs, build output) and removes the staged copy when the review finishes.

Remote HTTP endpoint (web / URL-based clients)

For MCP clients that connect to a URL instead of spawning a local process (Cursor remote servers, ChatGPT/Grok connectors, etc.), codeJung also runs a streamable-HTTP MCP server on the host — codejung_mcp_http.py. Unlike the stdio server, it runs on the codeJung host and talks to the API directly over localhost (no SSH). It exposes the PR tools only (review_pr, submit_review, get_review) — remote clients have no local dir to review.

Public endpoint: https://codejung.wint3rmute.com/mcp Auth: Authorization: Bearer <CODEJUNG_SERVICE_API_TOKEN> (enforced at the Caddy edge; the request never reaches the MCP server without it).

Exposure is off by default — open it for a demo with agentic_scripts/codejung-expose/codejung-expose.sh enable (and disable after).

Cursor (remote server)

"codejung-remote": {
  "url": "https://codejung.wint3rmute.com/mcp",
  "headers": { "Authorization": "Bearer <TOKEN>" }
}

Other web agents

ChatGPT (custom connectors) and Grok can point at the same URL, but their MCP client/auth support varies by product and plan (some expect OAuth rather than a static bearer header). The endpoint itself is a standards-compliant streamable-HTTP MCP server with a bearer gate; whether a given product accepts it depends on that product.

Host-side deployment (reference)

• venv: ~/.codejung-mcp-venv (pip install mcp)

• service: codejung-mcp-http.service (systemd), binds 127.0.0.1:8765

• Caddy: /mcp route, bearer-gated via CJ_MCP_TOKEN, rewrites upstream Host to 127.0.0.1:8765 (the MCP SDK's DNS-rebinding guard only trusts localhost)

Redeploy after editing this file: copy it to the host and sudo systemctl restart codejung-mcp-http.