ucp-mcp

# ucp-mcp — reference MCP server for Universal Context Packages Exposes UCP documents to any MCP-capable agent (Cursor, Claude Code, Codex, Gemini CLI, …). This is the reference composition of the two protocols: > **MCP is the pipe. UCP is what flows through it.** The server watches a directory of `*.ucp.json` files and serves them through three tools. Any producer (a context platform, a script, a CI job) can drop packages into that directory; any agent connected to the server can consume task context that is structured, provenance-backed, and token-budgeted. ## Install & run ```bash pip install ucp-mcp ucp-mcp --dir ./contexts # or: UCP_DIR=./contexts ucp-mcp ``` Cursor / Claude Code configuration (`mcp.json`): ```json { "mcpServers": { "ucp": { "command": "ucp-mcp", "args": ["--dir", "/path/to/contexts"] } } } ``` ## Tools | Tool | Purpose | |---|---| | `list_contexts()` | Inventory: entity id, title, system, freshness for every available package | | `get_context(entity)` | Full UCP JSON for an entity (by id, URL, or title fragment) | | `get_context_markdown(entity, token_budget?)` | Canonical CommonMark rendering (SPEC §7), optionally truncated to a token budget by salience | `entity` matching is forgiving: exact entity id (`PAY-482`), source URL, or a case-insensitive title fragment. ## Why a directory of files? This server is a *reference*, not a product. Its job is to demonstrate the MCP+UCP composition end to end with zero infrastructure, so that: - agent users can try UCP in one minute; - producers see the contract they need to implement (a real producer replaces the directory with a live Context Builder, keeping the same tools). ## Development ```bash pip install -e ".[dev]" pytest ```