MCP without the 10k-token cover charge

What it does

Pi MCP Adapter is a bridge between the Pi coding agent and standard MCP servers. Rather than dumping every tool schema from every connected server into the agent’s system prompt, it registers a single mcp proxy tool that searches and invokes capabilities on demand. It reads existing .mcp.json configs automatically and can adopt settings from Cursor, Claude Code, and Codex so you don’t rebuild your toolchain from scratch.

The interesting bit

The usual MCP tax is brutal: a single server can eat 10,000 tokens before you even say hello. This adapter shrinks that to about 200 tokens for the proxy, with servers staying disconnected until you actually call them. Tool metadata is cached locally, so searching and describing capabilities works even while the server is napping.

Key highlights

• One proxy tool (~200 tokens) replaces hundreds of individual tool definitions
• Lazy-by-default lifecycle: servers connect on first use and disconnect after idle timeout
• Imports existing configs from Cursor, Claude Code, Codex, or standard .mcp.json files
• Optional directTools mode promotes specific tools to the agent’s native list when you need them always visible
• Interactive /mcp panel for connection status, OAuth flows, and toggling direct vs. proxy registration

Caveats

• Newly configured directTools won’t appear immediately until the metadata cache warms up; they fall back to proxy on the first session
• MCP UI windows render natively only on macOS with Glimpse installed; otherwise they open in a browser tab

Verdict

Grab this if you’re using Pi and want MCP’s ecosystem without the context-window hangover. If you’ve already ditched MCP for hand-rolled CLI tools, keep walking.