MCP Tools

When running as an MCP server, cxp exposes four tools to your AI agent.

---

fetch_project_context

Discovers and summarizes new transcripts for the current project. Returns a compact index of all available summaries — ids, types, titles, and tags. Call this first to get the list, then use get_project_context to load specific ones.

When to call: At the start of a new conversation, or when the user references past work.

Parameter	Type	Required	Description
project_path	string	No	Absolute path to the project root. Defaults to cwd.

Returns: A markdown index listing all available summaries with ids, titles, types, and tags.

---

get_project_context

Loads the full markdown content of selected summaries into the agent's context window.

When to call: After fetch_project_context, when specific summaries look relevant.

Parameter	Type	Required	Description
project_path	string	No	Absolute path to the project root. Defaults to cwd.
ids	string[]	No	Summary ids from fetch_project_context. Omit to load all.

Returns: Full markdown content of the selected summaries.

---

search_context

Full-text search across all stored summaries. Searches both local ContextPool/ and any cached team insights.

When to call: When the user mentions a specific bug, error message, component, or decision. Before suggesting solutions to a problem — check if it was already solved.

Parameter	Type	Required	Description
query	string	Yes	Keyword or phrase to search for.
project_path	string	No	Limit search to one project. Omit to search all projects.

Returns: Matching insight excerpts with source file references.

---

list_context_projects

Lists all projects that have stored summaries, with insight counts.

When to call: When working across multiple repos, or when the user asks what's in memory.

(No parameters)

Returns: A list of project paths and their summary counts.

---

Agent Behavior Rules

The MCP server ships with built-in instructions that guide your agent to use these tools intelligently:

• First message in a new project → auto-call fetch_project_context

• User references a past conversation → search_context first, only fetch if nothing found

• Debugging an error → search_context with the error message or component name

• Making an architectural decision → search_context to check prior decisions

These instructions are embedded in the MCP server and don't require any configuration.