Soma MCP Server

Give your AI agent the one thing it can't do for itself: actually run code against tests and prove it passed.

Soma is an execution-verified code service. This MCP server exposes two tools:

• soma_verify_code — run candidate code against tests inside an isolated sandbox; get a PASS/FAIL verdict plus a signed, offline-checkable certificate (Ed25519). Use it to independently confirm code works before trusting it.
• soma_generate_verified_code — ask Soma to write code for a task; when the task is verifiable, the returned code has already been executed against derived tests, with a certificate attached.

15+ languages are supported for verification (Python, JavaScript/TypeScript, Go, C/C++, Java, Rust, Ruby, PHP, Bash, and more).

Install

Requires Node.js 18+. Runs over stdio.

Add to your MCP client config (Claude Desktop, Cursor, etc.):

{
  "mcpServers": {
    "soma": {
      "command": "npx",
      "args": ["-y", "soma-verify-mcp"],
      "env": {
        "SOMA_API_KEY": "YOUR_SOMA_KEY"
      }
    }
  }
}

• Claude Desktop: Settings → Developer → Edit Config, add the block above, restart.
• Cursor: Settings → MCP → Add, or drop the same block in ~/.cursor/mcp.json.

Configuration

Get a free preview key: contact centrum.arvind@gmail.com (free tier during the preview).

Tools

soma_verify_code

Run code against tests and return a signed verdict.

• language (string) — e.g. python, javascript, go, rust.
• code (string) — the complete source to verify.
• tests (array) — one of:
  • function mode (default): [{ "input": [arg1, arg2], "expected": value }] plus entrypoint (the function name).
  • stdio mode: set mode: "stdio" and [{ "stdin": "...", "expected_stdout": "..." }]; no entrypoint.
• entrypoint (string, optional) — function name for function mode.
• mode ("function" | "stdio", optional).

Returns: verdict, tests_passed, tests_total, and a signature / public_key / sig_alg you can check offline.

soma_generate_verified_code

Get code for a task, executed against derived tests before it's returned.

• prompt (string) — the coding task. Include concrete input/output examples (e.g. >>> f(2) == 4) so the result is verifiable rather than best-effort.
• max_tokens (int, optional, default 1500).

Returns: the code, certified (bool), and a certificate (verdict, tests_passed, tests_total) when verification passed. If a task isn't verifiable, output is returned uncertified and clearly labeled — never a false "verified".

What the certificate means

A certificate attests that the listed tests passed inside an isolated sandbox at generation time. It is signed (Ed25519) and checkable offline against the returned public key. It is not a warranty of fitness for any purpose — review output before production use.

Privacy

No training on your prompts. See the Soma Privacy & Data Policy at ${SOMA_BASE_URL}/privacy.

License

MIT.