neurodock-mcp-translation

Communication and translation tools as an MCP server. Shares its schemas and prompt library with @neurodock/extension-browser (Phase 2).

Version: 0.0.2 (developer preview, deterministic baseline).

Status

v0.0.x implements the four tools specified in ADR 0005 — translation tool design:

Design framing: deterministic baseline + optional LLM refinement

Per ADR 0005 §1, this server contains no LLM SDK (no anthropic, no openai, no ollama import). The substrate is provider-agnostic by construction. Each tool returns an envelope of the shape:

{
  "deterministic_analysis": {
    /* v0.1.0 output shape, populated heuristically */
  },
  "prompt_for_llm_refinement": {
    "role": "user",
    "content": "<rendered prompt template>",
    "output_schema_ref": "packages/mcp-translation/schemas/<tool>.schema.json"
  },
  "eval_corpus_slice": "packages/evals/corpora/translation/<slice>.jsonl"
}

The caller's MCP client (Claude Desktop, Claude Code, a custom MCP host) can:

1. Use the deterministic analysis alone — useful when no LLM is available, when latency matters, or when the user has not yet configured a provider. The deterministic baseline already detects common ambiguity patterns (circle back, let's revisit, next week), scores tone on simple word-list heuristics, and partitions meeting transcripts via speaker-prefix and regex matching.
2. Refine via its own LLM — feed prompt_for_llm_refinement.content to the caller's configured model (Claude, GPT, local Llama) and ask for a JSON response conforming to the schema at output_schema_ref. Replace the deterministic analysis with the refined object before rendering to the user.

This makes the server useful without a connected LLM and vendor-neutral when one is connected.

Verbatim-anchor enforcement (brief_meeting)

brief_meeting.ambiguous_items[].quoted_span.text MUST equal input.transcript[start_char:end_char]. The server validates this on every response (including its own deterministic output) and raises VERBATIM_ANCHOR_FAILED rather than fabricating ambiguity that is not in the transcript. This is anti-hallucination armour required by plan.md §7.

References

• Spec: Section 7.
• Tool design rationale: docs/decisions/0005-translation-tool-design.md.
• Authoritative schemas: packages/mcp-translation/schemas/.

Running

uv run neurodock-mcp-translation

The server speaks the MCP stdio transport.

Smoke test:

uv run python -c "from neurodock_mcp_translation import server; print(server.app.name)"
# neurodock-mcp-translation

Development

uv sync --all-packages --all-extras
uv run pytest packages/mcp-translation/tests/ -v
uv run ruff check packages/mcp-translation/
uv run ruff format --check packages/mcp-translation/
uv run mypy --strict packages/mcp-translation/src/

Known limitations (v0.0.2)

• No language packs. The deterministic heuristics are English-only. BCP-47 target_language is accepted but only passed through to the prompt.
• No model_provenance from a real LLM. The deterministic analysis sets mode="unknown" on every response; a connected LLM client is responsible for setting this in the refined response.
• No eval-corpus integration yet. The eval_corpus_slice paths are recorded on every response but the slices themselves are owned by eval-curator and land in Phase 2.
• No streaming. Long meeting transcripts (>200k chars) must be chunked by the caller.
• The four schemas under schemas/ are the v0.1.0 wire contract; the v0.0.2 envelope wraps them and will collapse to the wire contract once the Phase 2 LLM-refinement flow is the default path.