Skip to content

design.md

Latest commit

 

History

History
67 lines (43 loc) · 1.64 KB

design.md

File metadata and controls

67 lines (43 loc) · 1.64 KB

Design

Goal

Keep Claude Code as the cloud-side coding agent while moving noisy context-reading work to a local LLM.

Tagline:

Claude in the cloud. Eyes on the ground.

Separation of roles

Claude Code:

  • plans
  • edits
  • debugs
  • reviews
  • decides

Local reader:

  • reads long docs
  • compresses logs
  • summarizes diffs
  • ranks grep output
  • extracts API surfaces

The local model is not the main agent. That boundary is the core design choice.

Why wrappers instead of deeper integration

v1 uses simple shell wrappers because they are:

  • easy to inspect
  • easy to install
  • easy to remove
  • compatible with Claude Code's existing shell-tool workflow

The wrappers keep the integration obvious: a local read task goes through ask-local, and Claude Code remains in charge.

ask-local

ask-local is the only script that talks to LM Studio directly. The other commands are thin task-specific wrappers that shape the question and input:

  • qdoc for docs
  • qlog for logs
  • qdiff for diffs
  • qgrep for grep output
  • qapi for API extraction

This keeps the operational surface small and beginner-friendly.

Why OpenAI-compatible LM Studio

LM Studio exposes an OpenAI-compatible API on localhost, which makes it easy to reuse a standard Python client and avoid custom transport logic.

Non-goals

  • replacing Claude Code's main model
  • using ANTHROPIC_BASE_URL or ANTHROPIC_AUTH_TOKEN
  • automatic code editing by the local model
  • claiming the local model's summary is authoritative

Future direction

MCP-based integration could be explored later, but it is not part of v1. The first version stays intentionally small and shell-driven.