synapse-mcp: Instructions file for Claude Code

Project

FastMCP-based server that exposes Synapse (synapse.org) capabilities to LLMs over MCP. Tool catalog is 72 tools (48 read-only, 24 write/destructive) spanning entity, wiki, team, user, activity, evaluation, submission, schema, organization, form, curation, utility, and search domains.

Stack

• Python 3.11 (CI runs on 3.10)
• FastMCP 3.2.3 (fastmcp)
• synapseclient 4.12.0
• pytest with anyio (asyncio backend) — see tests/CLAUDE.md
• astral/uv for local dev

Commands

• Install deps: uv sync
• Run the full test suite: uv run pytest
• Run a single test file: uv run pytest tests/test_tool_service.py
• Run the BM25 eval with its printed summary: uv run pytest tests/evals/test_tool_selection.py -s
• Run the server over stdio: uv run synapse-mcp
• Run the server over HTTP on :9000 (matches the VS Code launch config): uv run synapse-mcp --http --host 127.0.0.1 --port 9000

Tool authoring

Every tool is declared via @service_tool (from synapse_mcp.services), never @mcp.tool directly — because the decorator enforces naming, docstring, and tagging conventions at registration time. See doc/tool-authoring.md for the full convention.

• Tool name must start with one of: get_ list_ search_ create_ update_ delete_ submit_ check_ validate_ register_ — decorator raises ValueError at import otherwise.
• First sentence of description= must name the concrete Synapse object passed as synapse_object= (head-noun matched, plural allowed) — because LLMs pick tools by reading that sentence first, and a typo'd object name makes selection accuracy drop.
• Synonyms live in synonyms=(...); siblings live in siblings=(...). Never embed them in prose — because the Related terms: / Distinct from: lines the decorator renders keep the primary description scannable for humans while still feeding the BM25 index.
• Tool function bodies stay thin: validate ID with validate_synapse_id(...), delegate to a service-class method, return — because business logic lives in services/ under @error_boundary. Any logic that escapes the tool wrapper becomes a 500 to the LLM.
• ID-accepting parameters must include a concrete example in the description: syn123456 for entity IDs, numeric strings for team/evaluation/submission/user IDs, URI form for JSON schema $id — because LLMs infer argument shape from example values, not just parameter names.
• Tool functions put ctx: Context last, EXCEPT when every business arg is optional (defaulted): then ctx comes first, because Python forbids a non-default arg after a defaulted one (e.g. get_entity_provenance(ctx, entity_id=None, ...)). Service methods take ctx: Context first, then business args — no self, since services are @staticmethod. Don't swap the service/tool convention.