ollama-code-mcp CLAUDE.md

Routing guidance: local Ollama vs. cloud Claude

This project exposes coding tools backed by a local/LAN Ollama instance (Qwen3) via MCP. Use this file to decide when to delegate a task to those tools versus doing it yourself.

Delegate to ollama-code-mcp when the task is...

• Mechanical or repetitive. Writing a first-draft test suite, generating boilerplate (CRUD handlers, DTOs, config parsing), or applying the same mechanical edit across many files (batch_refactor).
• Context-window-heavy but low-ambiguity. Explaining an unfamiliar file, reviewing a large diff for obvious issues (code_review_diff), or reviewing/refactoring a big file you'd otherwise have to paste in full. Use file_path / diff_file / glob_pattern instead of pasting content so the file never enters your own context window.
• A useful first pass before your own review. Run review_code or code_review_diff to surface candidate issues, then apply your own judgment to what it finds rather than trusting it blindly -- local models are meaningfully weaker than you at subtle correctness reasoning.
• Not security- or architecture-sensitive. Straightforward fixes, explanations, and tests for code where being wrong is cheap to catch and correct.

Keep in the cloud (do it yourself) when the task is...

• Architecturally significant -- API design, cross-module refactors that require holding the whole system in mind, anything where a subtly wrong answer is expensive to unwind.
• Security-sensitive -- auth, crypto, permission checks, anything touching secrets or user data. Don't delegate generation or fixes here; you can delegate a review pass as a second opinion, but verify its findings yourself.
• Ambiguous or under-specified. If the task needs clarifying questions or multi-turn negotiation with the user, handle it yourself -- the Ollama tools are single-shot, not conversational.
• Small and fast. A one-line fix or a quick rename isn't worth a network round-trip to another host; just make the edit.
• Already failing. If a tool call reports Ollama is unreachable, timed out, or the model isn't pulled, don't retry it -- handle the task yourself. Call ollama_status once to confirm before writing off local delegation for the rest of the session.