Local AI Toolkit

Using it with a cloud agent

Contents · Русская версия

Nothing else to set up: the installer has registered the MCP servers and told the agent when to use them. From there the agent decides — it sees a “where does this live” question and reaches for search_code instead of reading files.

What is worth knowing so it works fully rather than halfway:

Ask in words, not in names. “Where is it decided how long to wait before retrying” beats guessing a class name. Questions like that are exactly where an index beats grep.

Uncommitted edits are visible, but not by themselves. The index catches up on command: index_refresh from the agent, or localai sync by hand. After a commit the hook does it.

A refusal is not a breakage. “No overlay matches this worktree” means the tree moved and the index is behind; the answer is to build it, not to route around it.

Verify anything a decision rests on. A local model is wrong more often than a cloud one, and the product does not hide it: every answer is signed with its model and its duration.


← Installing · Contents · Using it from the command line →