Zero-model recall over your files. Ask a question, get back the exact section that answers it and the file it came from, in about a millisecond, without calling a model. Built for Claude Code: as an MCP server it replaces grep-and-read-whole-files with one call, using far fewer tokens.

The retrieval engine inside Laika Orbit, published on npm as laikaorbit.

Use it with Claude Code

claude mcp add laikaorbit -- npx -y laikaorbit mcp

Claude Code starts the server in your project folder and gets three tools: recall (answer a question with the matching section and its source), get (read a file) and status. To point it at another folder, add --root <dir> after mcp.

Use it from the terminal

npx laikaorbit index                          # build the index for this folder
npx laikaorbit recall "how do we roll back a deploy"
npx laikaorbit lint                           # check router files for dangling pointers
npx laikaorbit ask "…"                        # print the packed prompt for a model

Every command takes --root <dir>; otherwise the corpus is BRAIN_ROOT, then the working directory.

How it finds the answer

Seven steps, none of them a model call: tokenise the question on word boundaries, score it against an inverted index, keep the top five with a confidence margin, open only the winner, slice out the matching ## section, follow at most one pointer, and pack the result.

It works on any folder of Markdown and text. It works best with router files: short typed pointer lists in brain/routers/*.md that say where things are, written the way questions are asked:

## 1 — deploys

- Files: notes/deploy.md — how production deploys are rolled back
- Rules: notes/freeze.md — no deploys on Fridays after 2pm

Which folders are indexed, and how, is set in brain/index.config.json (optional). .docx, .rtf and .odt are read with macOS textutil, and PDFs with pdftotext when it's installed.