Tomosu MCP: govern AI-generated code without leaving your editor.

1. Connect the server

The hosted server needs no install. Point your client at https://mcp.tomosu.ai/mcp and it negotiates the rest. Pick the client you actually use:

Claude Code

claude mcp add --transport http tomosu https://mcp.tomosu.ai/mcp --scope user

Then run /mcp inside Claude Code to sign in. --scope user makes the server available in every project on this machine; drop it to add the server to the current project only.

Claude Desktop

Open Settings → Developer → Edit Config and add the server to claude_desktop_config.json:

{
  "mcpServers": {
    "tomosu": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.tomosu.ai/mcp"]
    }
  }
}

Quit and reopen Claude Desktop. The Tomosu tools appear under the tools icon in the composer.

Cursor and other MCP clients

Any client that speaks streamable HTTP can use the same URL. Where the client asks for a command rather than a URL, use the mcp-remote bridge shown above.

2. Authenticate

Every tool is scoped to your account, so nothing works until you sign in. In Claude Code, run:

/mcp

Pick the Tomosu server and follow the browser prompt. The server issues your token and your client stores it; other clients run the same OAuth flow on first tool call.

3. From install to PRI score

This is the whole loop. Steps 3 to 7 are one conversation with your assistant — you are not calling these tools by hand.

suggest_code(
  files=[{"filepath": "src/checkout.ts", "filecontent": "..."}],
  message="Analyze for reliability issues: crashes, race conditions, "
          "missing error handling, unhandled promise rejections",
  project_name="storefront",
  repo_path="https://github.com/acme/storefront",
  context=["reliability", "error", "data_integrity"]
)

save_recommendation(
  scan_uuid="…",
  accepted=True,
  filename="src/checkout.ts",
  before_code="…",
  after_code="…",
  project_name="storefront",
  severity="high"
)

get_scan_report(
  scan_uuid="…",
  repo_path="https://github.com/acme/storefront"
)

curl -L "{returned_url}" -o tomosu-report.pdf

get_dashboard_analytics(project_name="storefront")

Tool reference

Seven tools, all scoped to the authenticated account.

Analysis contexts

Pass one to three of these in context to steer suggest_code. For the reliability workflow, ["reliability", "error", "data_integrity"] is the useful default.

Prompt templates

The server also ships three prompts. They take code and an optional filepath, and drive suggest_code with the right contexts already chosen.

Run it yourself

The hosted endpoint suits most teams. Run your own when code must not leave your network.

With Docker Compose

The server sits alongside the python, node and agent backends and listens on 8080:

docker compose up -d

Then point Claude Desktop at your instance:

{
  "mcpServers": {
    "tomosu": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "http://localhost:8080/sse"]
    }
  }
}

From source, over stdio

Requires Python 3.10+ and uv.

cd ai-dev/mcp-server
uv sync
uv run server.py

Exposing a local server to a remote client

To reach a local instance from a client that needs a public URL, tunnel port 8080 and set MCP_ISSUER_URL to the generated address — OAuth breaks if the issuer does not match what the client connected to. Then point the client at <your-url>/mcp.

brew install cloudflared
cloudflared tunnel --url http://localhost:8080

Environment

Testing the tools directly

npx @modelcontextprotocol/inspector uv run server.py

Troubleshooting