A governed route
Local models by default, free lanes within budgets and breakers, frontier spend capped at $0 unless raised.
Home / Karta · In internal evaluation — not released
An agent inside your limits.
Karta is the AI coding agent we are building for teams that need control: it works inside the permissions a task allows, on a cost-governed route, stops for a person where it should, and keeps a record of every step.
01 / 08
The route
Before any work, a request takes a governed route: our own GPU mesh first, in an order people set and checked for health; vetted free hosted lanes only if the mesh is down, each checked for credentials, remaining daily budget and a tripped breaker; frontier models behind a gate whose spend cap is $0 by default. It is a governed order with live checks, not a live auction for the cheapest answer.
02 / 08
The grant
Karta sees the repository as a city of files. A task comes with a grant — the parts it may change — and a fence round protected paths: deployment, its own safety tooling, the rules it works by.
03 / 08
The work
It reads, plans and proposes a change as a diff — here, adding input validation to a form handler in a sample repository (a fictional task).
04 / 08
Approval and tests
Touching a protected path stops the work until a person approves; the approval fails closed and cannot be replayed. Then the project's own checks and tests decide whether it is done — a failing test sends it back to revise.
05 / 08
The trace
Every step is written to a record. Older steps are compacted, yet each can be expanded back into its detail when needed. Security-relevant events go to a separate, hash-linked chain that can be verified end to end.
06 / 08
Context and memory
Long histories fold without losing their errors — failure lines are always kept. What Karta learns is stored as a lesson only after the task that produced it passes its gates.
07 / 08
Improving, behind a locked door
Karta can propose improvements to itself, built apart from the running version: a test that fails before and passes after, a person's approval, a shadow run and a way back. Changes to its safety controls are never promoted on their own, and nothing here runs unattended In development
08 / 08
The control room
A console will show it all in one place — health, budgets, approvals waiting, the verified record. Today the only interface is a terminal; the web app is in development In development
Control, not autonomy.
The parts running in our internal testing today.
Permissions and protected paths
A grant for each task; protected paths it cannot change without a person.
Approvals that fail closed
A pending approval is bound to what it approves and cannot be used twice.
Tests decide “done”
Linting, type checks, the project's tests and a safety suite it cannot edit.
A record you can verify
Every step recorded; security events in a hash-linked chain that detects tampering.
Sandboxed work
Commands run isolated, with the network off and credentials kept out, by default.
The route, stated plainly.
| Step | How it decides |
|---|---|
| Our GPU mesh | First, always: models in an order people set, each checked for health; the first healthy one runs |
| Free hosted lanes | Only when the mesh cannot serve: a vetted order, then live checks of credentials, remaining daily budget and circuit breaker |
| Frontier models | Behind a gate with a monthly spend cap of $0 by default |
| Learning the best route from data | In development |
| A running record of spend | In development |
A task, its grant and its record.
Illustrative: the checks a Karta task passes through, composed from the controls running in our internal testing. The task is fictional.
- Task
- Add input validation to a form handler
- Grant
- The parts of the repository this task may change
- Blocked
- An edit to a protected path waits for a person's approvalThe approval is bound to that edit and cannot be used twice
- Checks
- Linting, type checks, the project's tests and a safety suite Karta cannot edit
- Record
- Every step written; the security chain verified end to endVerification reports whether, and where, the chain breaks
What you can engage us for today
- Conversations about governed AI engineering, and walk-throughs of Karta's controls.
In development
- The Karta web console: health, budgets, approvals and the record.
- Routing learned from data rather than a set order.
- A release: Karta is evaluated internally and there is nothing to install yet.
Talk to us about Karta.
Tell us what an AI agent would need to be allowed to touch in your repositories, and what it must never touch. We will show you how Karta keeps to that.