Three things, and nothing else.

Latorium registers as an MCP server. Your agent keeps its own model, its own loop, and its own judgment. What changes is how much it has to read — and what happens once you have nothing left to spend.

Install the free tier Privacy

It retrieves, so your agent doesn't.

Claude Code and Codex call two tools instead of opening files directly.

local_search returns matching excerpts from the workspace. context_packet builds a bounded, task-relevant brief. Both run a local model over a local index, and both hand back passages rather than files.

Each call takes an agent argument, so the dashboard can attribute what it avoided to the agent that asked. Calls that don't identify themselves are reported separately rather than guessed at.

It writes, when you've nothing left to write with.

Summaries, test scaffolding, small refactors, boilerplate — handed to ollama_draft on your own hardware.

Day to day this is just cheaper: work that never needed a frontier model doesn't consume one. The reason it matters is what happens when your usage window is spent, or you're offline, or the network is gone. The local model doesn't have a window. It carries on at whatever standard it can — which for single-file work is about seven times in ten, and which you can read on the home page rather than take on trust.

The output is a draft, and it is treated as one. Nothing is applied without your review, and the extension will not write a file you haven't approved.

ollama_draft(prompt: "unit tests for retry escalation")
  drafted 41 lines locally
  awaiting review — no files written

The Latorium sidebar in VS Code, showing Ollama ready, qwen3:8b ready and folder open, a performance selector, and prompts for understanding a project, fixing an error, building a feature or reviewing code. — The panel checks Ollama, the model and the workspace before it offers to do anything, and says which of the three is missing when one is.

It reports what it kept out, per agent.

Claude and Codex are measured separately, against their own session logs read from disk.

Only token counters are read — never prompt or response content. Cached re-reads are excluded from the denominator, because they re-present context the session already took in; counting them would shrink the figure as a conversation got longer rather than as retrieval got worse.

Claude, newly ingested context: read from session log
Codex, newly ingested context: read from session log
Attributed to neither: reported separately

The dashboard ships with nothing filled in. It fills itself in from your machine, which is the only place the honest number exists.

External-agent context in the dashboard: MCP tokens avoided, context packets produced, and Codex input and output counters read from the local session log. — The lower half is per-agent accounting. Codex figures come from its own session log on disk; Claude's arrive the same way.

Questions

Does this replace Claude Code or Codex?

No. Latorium is a token saver first — it makes the agent you already pay for go further by doing the mechanical work around it. The local model can also keep coding once your usage window is spent, which is a useful fallback rather than a replacement: it will not out-code a frontier model, and the benchmark figures on the home page say precisely how far it gets.

What do I need installed?

VS Code, Ollama, and a local model of your choosing. Latorium doesn't ship a model. Any MCP-capable agent can use the server — not only Claude Code and Codex.

Can I run it air-gapped?

Yes. Once Ollama and your model are staged, Latorium needs no internet at all. Licence verification is a local signature check, not a server call.

What happens when a licence expires?

You drop back to the free tier. Nothing is deleted and nothing is locked — your task board and saved lessons stay where they are.