How Latorium works

Latorium runs as an MCP server beside your coding agent. When the agent asks for repository context, Latorium searches locally and returns a focused packet. For smaller jobs, Local Coder can draft the work on your machine.

Install the free tier Privacy

Start with the request you already have

Tell Claude Code or Codex what you want: “add dark mode”, “fix this login error”, or “make the button save my form”. You do not need to know where the relevant files are.

You ask normally. Keep using the coding agent you already use.
Latorium finds the useful code. It searches on your computer and picks the pieces related to your request.
Your agent gets a focused packet. It receives the relevant code, file names, and connections.
The agent makes the change. Latorium can flag risky files and suggest what to test.
You can see what happened. The dashboards show what was chosen and how much context was avoided.

MCP is the connection between the agent and Latorium. Set it up once; you do not need to type MCP commands yourself.

Latorium finds the relevant code first

Your agent asks Latorium for the parts of the workspace related to the current job.

local_search returns matching excerpts from the workspace. context_packet builds a brief with a fixed size limit. Both return relevant passages with their file paths.

Each call takes an agent argument, so the dashboard can attribute what it avoided to the agent that asked. Unidentified calls appear in a separate total.

Advanced tool reference

Name	Plain-English job	Example
`local_search`	Find the right code without opening every file.	Where is checkout calculated?
`context_packet`	Make one compact cheat sheet for a task.	Everything relevant to adding coupons.
`context_review`	Catch an old copy of a file.	Replace yesterday's auth code.
`impact_of_change`	Show what might break before editing.	This type is used by four services and two tests.
`local_coder_draft`	Give routine writing to Local Coder.	Draft boilerplate or tests for review.
`create_task` / `claim_task`	Split work without agents colliding.	One owns the UI; another owns the API.

Then it handles smaller jobs locally

For focused work such as summaries, tests, boilerplate, and small refactors, local_coder_draft can use the model running on your own machine.

This keeps routine work off paid model usage. Local Coder can also run when the network or a provider is unavailable. Results depend on the task and hardware; the benchmark page shows the measurements.

The output is a draft. Nothing is applied without your review, and the extension will not write a file you have not approved.

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

The Latorium sidebar in VS Code: a Ready indicator, three cards explaining that code stays on this computer and that retrieval and local drafting cut input and output tokens, toggles for Focus on code and Local AI, a Balanced 6K response mode, starting prompts, and a live usage and savings panel. — The panel checks Local Coder, its model and the workspace before it offers to do anything, and says which of the three is missing when one is. Usage and savings update in the same panel, so the figure is next to the thing producing it.

See the context it kept out

Claude and Codex are tracked separately using the token counters already available in their local session logs.

Latorium reads token counters from local session logs. It does not read prompt or response text. Cached input is reported separately from newly read input.

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

The dashboard starts empty and fills in as Latorium is used on the machine.

The Latorium overview: project reading avoided, how much smaller requests were, a before-and-after token comparison, writing assistance broken down by kind of work, local AI activity, per-agent reduction for Claude and Codex, MCP retrieval totals, and verified outcomes. — Per-agent accounting sits partway down. Codex figures come from its own session log on disk; Claude's arrive the same way. Kinds of work whose measured yield falls below fifteen percent are refused from then on, which is why one row reads as refused rather than simply low.

Release checks catch problems before customers do

The current development build runs adversarial and benchmark checks against the same retrieval and agent loop used in the extension.

A release must pass the test suite, 20 adversarial scenarios, TypeScript checks, compilation, packaging, secret scanning, and regression checks.

Current development build: 746/747 of Latorium's own tests passed (one skipped, it needs symlink privileges) and 20/20 adversarial checks passed. Checked 19 September 2026.

Questions

Does this replace Claude Code or Codex?

No. Latorium reduces the context your existing agent needs. It handles surrounding mechanical work, while Local Coder can continue with focused tasks after a usage window is spent. Larger jobs may still need a stronger model.

What do I need installed?

VS Code and enough free disk space for Latorium Local Coder. Latorium's guided setup installs the required Local Coder runtime and model for you. Any MCP-capable agent can use the server, not only Claude Code and Codex.

Can I run it air-gapped?

Latorium Local Coder can continue without internet after setup. Only account refresh needs a connection.

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.