Turn MCP usage into learned metadata

AdaptiveMCP is a learning layer over Model Context Protocol (MCP) primitives. It observes how MCP tools are actually used and attaches learned metadata to existing primitives — it does not replace MCP, redefine tools, or add protocol abstractions. Clients adapt from real signal instead of guessing.

Status: experimental. The packages are published, but the API may shift before 1.0.

• 🧩 A learning layer, not a new protocol — observes tool usage and attaches learned metadata to existing MCP primitives.
• 📊 The store is authoritative — all metadata lives in a SQLite store (node:sqlite); the tools-metadata.yaml view is a derived projection, recomputed whenever metadata changes — never edited by hand.
• 🔌 Server governs, client executes — the server publishes a narrow, server-governed resource (dev.adaptivemcp/tools-metadata); the client reads it, adapts, and reports observations back, degrading gracefully on any host that ignores it.
• 💪 Modular, dependency-light packages — all @adaptivemcp/* libraries are published to npm.

Data always flows in one direction: event → MemoryStore → derived YAML view.

Package	Version	Description
@adaptivemcp/spec		Extension identifiers, event schemas, and shared types for Adaptive MCP.
@adaptivemcp/graph-analysis		Graph analysis for Adaptive MCP execution DAGs: critical paths, bottlenecks, failure cascades, anti-patterns, and workflow forecasting.
@adaptivemcp/memory		Persistent operational knowledge for Adaptive MCP, backed by SQLite (node:sqlite).
@adaptivemcp/telemetry		Tool execution events and observability for Adaptive MCP.
@adaptivemcp/evaluation		Outcome scoring and feedback loops for Adaptive MCP.
@adaptivemcp/extension		Adaptive MCP extension: derives the YAML tools-metadata view from the SQLite store and serves it to MCP clients.
@adaptivemcp/runtime		Batteries-included Adaptive MCP runtime: wires telemetry, evaluation, routing, orchestration, approval, and the extension into one transport-agnostic loop.
@adaptivemcp/routing		Model selection and cost optimization for Adaptive MCP.
@adaptivemcp/orchestration		Composition and execution strategies for Adaptive MCP.
@adaptivemcp/approval		Intent, plan, and tool approval boundaries for Adaptive MCP.
@adaptivemcp/thin-client		Minimal client-side execution loop for Adaptive MCP.
@adaptivemcp/middleware		Pluggable middleware chain for Adaptive MCP (Transform I/O, gate, inject-auth, observe).
@adaptivemcp/mcp-binary		Generic CLI-binary -> MCP-server stdio wrapper (the only sanctioned shell-out layer).
@adaptivemcp/opencode-plugin		OpenCode plugin for Adaptive MCP - maps OpenCode hooks to Adaptive MCP telemetry and middleware

The packages above are the core. They are exercised and documented by sibling projects:

• 📦 adaptive-mcp — the monorepo with the @adaptivemcp/* sources, runnable examples, and the dev.adaptivemcp/tools-metadata SEP draft.
• 🤖 agent — an MCP-native agent host (@adaptivemcp/agent) that owns both the LLM completion and the tool execution, so the learning loop runs against a real client rather than only synthetic demos.
• 🧹 qc — QueryCleaner, a real third-party MCP server (Python) that publishes dev.adaptivemcp/tools-metadata; the agent drives it end to end in CI.
• 📚 adaptivemcp.github.io — the documentation site, adaptivemcp.dev.

• 📚 Docs & examples: adaptivemcp.dev
• 🚀 Getting started: adaptivemcp.dev/guide/getting-started
• 📦 Packages: adaptivemcp.dev/packages
• 🧪 Examples: adaptivemcp.dev/examples

We're just getting started 💪