lokomotiv.dev
Pre-release · private build

Set a goal. Wake up to verified work.

lokomotiv.dev is a local-first autonomous development factory. Give it one bounded goal. It plans the work, runs your own coding agent in an isolated Git worktree, has an independent reviewer check every change, runs your checks, and records evidence you can verify.

The factory

A crew of agents on your machine, not a chat window.

Works with Claude Code, Codex, Cursor Agent, or a custom command as the writer. Planning and review run on separate models. It asks you only for decisions your policy does not grant.

A finished goal open beside its crew: Plan, Build, and Check stages on a canvas, a passed attempt, and verified evidence
A finished goal beside the crew that ran it. All screenshots use fixture data.

How it works

Four steps from intent to evidence.

01

Set a goal

Describe one bounded outcome: what done means, which files may change, and how many attempts each task gets. An unattended goal adds a budget and keeps advancing within the limits you approve. A goal an agent proposes starts only when you choose to start it.

The Writer agent proposing a bounded goal with Start goal and Edit buttons
An agent proposes a goal: outcome, done-when checks, files, attempts, and budget.
02

A crew does the work

A planner writes a specification and task graph that a plan reviewer on a different model must approve. For each task the writer works in its own Git worktree, an independent reviewer checks the change, findings get at most one repair, and your verification commands run.

The command palette with Set a goal, Create a manual task, and Pause scheduling
The command palette: set a goal or jump to any page.
03

Decide when it asks

It asks when policy does not cover the next step, such as a failed verification or a reached limit, and it stops rather than guesses. A decision offers only the options the server presents, needs a reason, and is rechecked before it applies.

A decision after failed verification: retry or cancel, with a reason
A decision: only the options the server offers, and a reason for the record.
04

Read the evidence

A result says Verified only when its evidence rebuilds to the recorded hashes; otherwise the claim is withheld. Each run keeps its diffs, logs, a SHA-256 manifest, a lineage record, and a hash-chained event log inside the repository. Merging stays with you.

A verified result: what changed, how it was checked, review, and delivery
A verified result: what changed, how it was checked, the review, and how far delivery went.

Hard limits

What lokomotiv.dev will never do.

These limits are enforced in code, not settings. No configuration or prompt removes them.

  • Merge, deploy, roll back, or force-pushDelivery can only commit locally, push a normal update to its own integration branch, and open or update a pull request.
  • Deliver an unverified candidateIt commits, pushes, or opens a pull request only when policy, repository authorization, and the goal's delivery limit all allow it, and only for the exact candidate that passed review and verification.
  • Send your code to a model without consentIts own planning and review reach a remote endpoint only with recorded consent for that provider, model, and purpose. Prompts are secret-redacted first.
  • Call a result verified when it is notIf the evidence no longer matches its hashes, the claim is withheld.

These limits cover lokomotiv.dev's own actions. Your writer agent talks to its own vendor directly, and the default native runtime is a Git worktree, not a sandbox. An optional container runtime (Docker or Podman) gives each command one workspace mount, no network by default, and scoped credentials.

Under the hood

Local-first, loopback-only, built to run overnight.

Core
GoA single CLI and service: scheduler, policy engine, worktree isolation, and evidence store. Runs as a LaunchAgent on macOS or a systemd user unit on Linux.
Dashboard
React and TypeScriptEmbedded in the binary and served only on loopback behind a session token.
Desktop
Tauri and RustA native macOS app around the same dashboard, built from source today.
Evidence
SHA-256 manifestsDiffs, logs, lineage, and a hash-chained event log, kept in the repository under .lokomotiv/runs/.
Models
Bring your ownAnthropic, OpenRouter, Ollama, any OpenAI-compatible endpoint, or a Claude or ChatGPT subscription CLI.
Hosts
macOS and LinuxApple Silicon is supported. Linux amd64 and arm64 are supported on server-style hosts.

Status

Pre-release. No public build yet.

The first lokomotiv.dev release, 0.3.0, is not cut yet. The source is private.

  • Works todayUnder the installed service, an unattended goal keeps advancing after the browser closes, and it finishes only when its recorded evidence verifies.
  • NextThe CLI, service, and embedded dashboard as a signed private release.
  • After thatThe macOS app, once it has Apple Developer ID signing and notarization.