What problem does this solve?
Small local models often lack knowledge of lmcode itself, so they give inaccurate answers to meta-questions like "what commands can I use?" or "what tools do you have?" — relying on training data rather than live state.
Proposed solution
Add an internal MCP (Model Context Protocol) server that exposes lmcode's own capabilities — slash commands, registered tools, current config, and session state — as MCP resources and tools.
This is a separate internal MCP server from the public FastMCP integration discussed elsewhere. It is not user-facing; it runs in-process and is only accessible to the agent itself.
Resources to expose:
lmcode://commands — list of all slash commands with descriptionslmcode://tools — registered tools with their signatures and docstringslmcode://config — current settings (model, mode, verbose, etc.)lmcode://session — turn count, context tokens used, active file
Target version: Pin to the same milestone as the broader MCP / agent-mode work (v0.4.0 or later).
Alternatives considered
None considered.
What problem does this solve?
Small local models often lack knowledge of lmcode itself, so they give inaccurate answers to meta-questions like "what commands can I use?" or "what tools do you have?" — relying on training data rather than live state.
Proposed solution
Add an internal MCP (Model Context Protocol) server that exposes lmcode's own capabilities — slash commands, registered tools, current config, and session state — as MCP resources and tools.
This is a separate internal MCP server from the public FastMCP integration discussed elsewhere. It is not user-facing; it runs in-process and is only accessible to the agent itself.
Resources to expose:
lmcode://commands— list of all slash commands with descriptionslmcode://tools— registered tools with their signatures and docstringslmcode://config— current settings (model, mode, verbose, etc.)lmcode://session— turn count, context tokens used, active fileTarget version: Pin to the same milestone as the broader MCP / agent-mode work (v0.4.0 or later).
Alternatives considered
None considered.