MCP surface expansion: cross-provider context graph

[whaleen]

Vision

Expand px mcp serve so any agent can connect and traverse the full cross-provider context graph for a project. Today the MCP surface is thin (project inspection only). The goal is: an agent starting a session calls project-index and immediately knows everything every other agent has learned about the project.

Proposed tools

Tool	Returns
px_sessions	All sessions across Claude, Codex, Gemini, Pi, Grok — with inline summaries and timestamps
px_memories	All memories for this project from all providers
px_skills	All installed skills — shared (~/.agents/skills/) and per-provider
px_context_files	All observed context files: CLAUDE.md, AGENTS.md, GEMINI.md, SPEC.md, .mcp.json etc. with content
px_providers	Which agents have been active in this project, last session time

Why this matters

• A Gemini session can discover what Claude worked on last week
• A new agent session gets full project context without the human having to re-explain
• Cross-provider knowledge is surfaced without any agent changing how it stores things
• project-index stays read-only — agents trust it won't mutate anything

Design notes

• Builds directly on the existing px mcp serve stdio transport — no new infrastructure
• All tools should accept an optional path argument (default: cwd) consistent with existing tools
• Output should be structured JSON suitable for agent consumption, not human-formatted text
• Consider a single px_project_context tool that returns everything in one call for agents that want a full snapshot

Prerequisites

• Complete provider compatibility table (issue Research: complete provider compatibility table #5) — need verified storage paths before wiring readers
• Grok Build support (issue Add Grok Build (xAI) support #4)

References

• Current MCP tools: px_project_inspect, px_agent_instructions, px_setup_plan
• cli/src/cli.rs — MCP server implementation
• docs/provider-table.md — compatibility matrix