mcp

Working with MCP (Model Context Protocol) servers in Freebuff Desktop: discovering tools via native search_mcp_tools / call_mcp_tool, bridging stdio servers via _scripts/mcp-call.py, injecting .env se

SKILL.md

GitHub

mcp — MCP tooling router

Home branch in Memora: root/skill/mcp.

Tool priority in Freebuff Desktop: (1) native search_mcp_tools + call_mcp_tool, (2) python _scripts/mcp-call.py <server> <tool> '<json>' when a stdio server is not yet mounted in the live session.

Choosing a part

Situation Part
Discover and call MCP tools in the current session parts/mcp-usage/
Batch install/repair servers, fix native bindings and env parts/mcp-batch-install-and-repair/
Configure .env secrets via the root .env exporter parts/mcp-auth-handler/
Diagnose tool-call failures, connection closed, moved paths parts/mcp-repair/
Build a new MCP server from scratch or from an OpenAPI spec parts/mcp-builder/

Rules

  • Walk root/skill/mcp in Memora before changing server configs; record probe results as nodes.
  • Use ask_questions before enabling/disabling servers in ~/.agents/mcp.json or adding new third-party dependencies.
  • Never inline plaintext tokens in ~/.agents/mcp.json — place secrets in the root .env and wrap the server command with the root .env exporter.
  • Pin package versions in ~/.agents/mcp.json (no blind @latest). On-demand servers live in _scripts/mcp-catalog-disabled.json.
  • If read_files returns [BLOCKED] on an ordinary non-secret file covered by .gitignore (such as inbox/ or _archive/), read it via the terminal (python / sed); never bypass secret rules for .env.

Related

memory (memora MCP), research (searchmcp and parallel MCPs), browser (browsermcp and playwright MCPs).