mcpgraphy 0.1.0

First release. pip install mcpgraphy.

A proxy for any MCP server that learns, from your agent's own calls, which tools work in which state, and warns or blocks before the next bad call. Every read, action result and refusal comes back with one line from the graph: the state, what works from it, what doesn't, and the shortest observed route to each other state.

Measured with a real GPT-5.4-mini agent on 200 jobs, start of job to end of job: jobs finished 74% → 99%, false success claims 10 → 0, rejected calls per job 1.03 → 0.43, tokens per job done 5,154 → 4,799. Full tables, the configurations tried on the way, and the command to rerun it are in docs/BENCHMARK.md.

Highlights:

• mcpgraphy wrap with observe / warn / block policies, --approve, --no-hints
• adapter JSON per server; adapters for the official filesystem server and two bundled simulators
• behavior graph with four edge statuses, journal, SQLite store, JSON and Mermaid export, drift detection
• research mode: simulators, safe explorer, action-offer MCP server, deterministic benchmark
• mcpgraphy benchmark llm: the real-agent benchmark with per-turn token attribution
• 101 tests, CI on Python 3.11 and 3.13