Your coding agent should not have to read the whole repo.

Latorium runs as an MCP server. Instead of opening files directly, your agent receives a bounded set of excerpts. It does not need to ingest the rest of the repository for that request.

Only the excerpts travel upstream. The Saved dashboard reports Claude and Codex separately, read from the session logs those tools already write — token counters only, never prompt or response content.

Keyword search works when you know the identifier. It is less useful when you describe a problem in plain language. Better ranking makes a packet more likely to contain the files an agent needs.

Right file in the top 10

63%100%

Ranking quality (MRR)

0.3780.598

Top hit, no shared vocabulary

0%50%

Keyword scoring against semantic ranking across 16 held-out questions on this repository; npm run retrieval-eval reproduces it. Semantic ranking is optional and off until you set a model — keyword remains the fallback, so nothing depends on a download.

Latorium can suggest the next small piece of code from the file and nearby symbols already in view, generated on your computer.

$ users.ts · activeUsers(users: User[])

→ latorium inline completion
  return users.filter(user => user.active);
  Tab to accept · 0 tokens upstream

  generated on this computer

Suggestions expire quickly, invalidate when source changes, and never queue ahead of a request you are waiting on.

Retrieval is the part you can see. The part that compounds is underneath it: Local Coder is a tuned adapter, trained on escalation traces — the cases where the local tier handed work upstream and a frontier model corrected it.

Those corrections are not all worth the same, so they are not weighted the same. Each lesson carries the strength of the model that produced it, and the adapter learns preferentially from the best teacher available at the time. As Claude and Codex improve, the thing running on your machine improves with them, without you changing anything.

This is why the paid tier is worth buying. Not a larger allowance — a model that has been corrected more times than the one in the free tier.

Traces are collected with consent, redacted before they are stored, and never include source, prompts or credentials. Correctness of the local tier alone is still being measured against real repository test suites; nothing on this page claims it passes them unaided.

A local 8B model does not replace a frontier model. Latorium handles search, focused reading, and routine drafting locally so Claude Code and Codex can spend their context on the harder work.

Your source, prompts, context packets, embeddings, and local-model output stay on your machine. Licence signatures are verified offline. If a licence lapses, the extension returns to Free; nothing is deleted or locked.

Everything it does