
mcp-toolrank-gateway
io.github.protortyp/mcp-toolrank-gateway
Search configured MCP upstreams and route explicit, allowlisted tool calls.
Documentation
MCP ToolRank
Discover, rank, and explicitly invoke allowlisted stdio/Streamable HTTP MCP tools. Model-free BM25 or hybrid with a verified 34 MB CPU encoder.
On 11 simulated workflows, the ToolRank hybrid cut schema tokens by 85% compared with giving Qwen all tools. Both dense models and the ToolRank hybrid completed 16/22 trials. Qwen completed 12/22 with all tools. Evaluation
Try it
The demo passes a clock reading between two local servers. It runs without a model download.
uvx --python 3.12 --from 'toolrank-cpu[gateway] @ https://github.com/protortyp/mcp-toolrank/releases/download/cpu-v0.6.0/toolrank_cpu-0.6.0-py3-none-any.whl' toolrank-gateway --demo
Import servers, inventory tools, and grant discovery and invocation separately. Search live from the shell. Pass the search_tools ID and revision to invoke_tool. CLI search · Import and inventory · Configuration · Gateway guide · CPU release · Desktop bundle
The CPU wheel ranks static tool catalogs with toolrank-cpu on Python 3.11–3.14. CLI and offline setup · OpenAPI import
Results and limits
ToolRank’s gain over BGE Small varies by ETOM schema under identical TF-IDF and BM25 scoring. Benchmarks and reproduction
Uncalibrated rankings leave applicability unverified. Invoke only tools permitted by the client and gateway. Calibration · Model, provenance, and license
LangChain and Pydantic AI examples · Browser demo · Research history · Apache-2.0
https://github.com/protortyp/mcp-toolrank/releases/download/gateway-mcpb-v0.3.0/mcp-toolrank-gateway-0.3.0.mcpb# mcpb: https://github.com/protortyp/mcp-toolrank/releases/download/gateway-mcpb-v0.3.0/mcp-toolrank-gateway-0.3.0.mcpb