The task outlives the chat.

A local control plane for AI coding agents. The brief, the limits, the open questions and the final decision live in one durable record on your machine — not in a transcript that ends with the session.

Install and run See how it works

Free and open source · Apache-2.0 · Runs entirely on your machine

Chat is a poor place to keep a task.

A session ends, a process dies, a provider changes. Anything that lived only in the transcript goes with it. Loomrail moves the parts that matter out of the conversation and into a record that a deterministic state machine owns.

Durable by default

State, requests, budgets, evidence and decisions are written to local SQLite in one transaction. Restart the daemon and the task is exactly where you left it.

Bounded by design

Every task carries allowed actions and a spend limit. The agent works inside them, and raising a limit is an explicit decision that gets recorded.

You stay the owner

An agent never closes its own task. Loomrail never commits, pushes, merges or deploys. The delivery is accepted or returned by a person.

How a task moves through Loomrail

Four steps, the same every time. New sessions run through your signed-in local Codex or Claude Code CLI; no API key is passed to Loomrail.

  1. Write the task, not the prompt

    A task states its outcome, its acceptance criteria, which actions the agent may take and how much it may spend. That contract is fixed before any work starts, and it is what the result is judged against later.

    CDC8DEDD Ready

    Persisted board integration

    Verify project isolation, state transitions and activity through the authenticated local API.

    Acceptance
    3 criteria
    Allowed actions
    Read, edit, run tests
    Budget
    Bounded and reconciled

  2. The agent works inside the limits

    Loomrail starts the provider session, watches the spend and advances the workflow state itself. The provider's output is treated as input to that state machine, never as the source of truth about progress.

    Workflow state

    1. Backlog
    2. Ready
    3. Running
    4. Review
    5. Accepted

    Budget guard Session limits + ledger

    Provider
    Local Codex or Claude
    Session
    Restarted once, state kept

  3. It stops and asks when it must

    When a decision needs a person, the agent raises a Human Request. The task blocks, the question stays attached to it, and your answer is recorded next to the work instead of scrolling away in a chat.

    Blocked Human Request

    Should the migration drop the legacy sessions table, or keep it read-only until the next release?

    Answer Return to work

  4. You accept the delivery, or send it back

    The task arrives with its evidence: the changes, the review, the QA result and any follow-up work it created. You inspect it and decide. Nothing is committed, pushed or merged for you.

    Evidence

    • Changes Snapshot-bound diff
    • Review Current report attached
    • QA Measured browser report attached
    Accept delivery Return to work

Try Loomrail without giving it a repository.

The copy block makes the package and Chromium downloads explicit, then checks your local Codex or Claude Code CLI on loopback. Sign in through that CLI first and use a new empty directory.

  1. Run the five visible commands. npm fetches Loomrail; Playwright fetches Chromium separately.
  2. The read-only preflight runs first. When it is ready, /try opens and names local state and log creation.
  3. Prepare the demo workspace, choose Codex CLI or Claude Code CLI, create the seeded task, move it to Ready and start the workflow.
  4. Answer the Human Request, approve a bounded provider budget if asked, inspect the available evidence, then decide yourself.

Loomrail uses the local CLI's existing subscription login. A missing, incompatible or signed-out CLI blocks dispatch; Loomrail never asks for a provider API key.

Runtime
Node.js 24.19–24.x
Network
127.0.0.1 only
First run
Local CLI preflight
Platforms
macOS Apple Silicon

What this macOS Stable release actually does

This release is Stable for macOS on Apple Silicon. Windows and Linux remain outside its support target and fail closed.

Available today

  • Same-machine browser UI, loopback daemon, and local SQLite state.
  • Codex CLI and Claude Code CLI adapters with fail-closed version/login checks and bounded workspace tools.
  • Restart recovery, Human Requests, budgets, evidence, and Decisions.
  • Repository registration, per-task worktrees, and change inspection.

Not claimed yet

  • Cloud sync, remote access, mobile control, or team accounts.
  • Windows/Linux live-provider support, automatic updates, or a desktop installer.
  • Automatic commit, push, merge, deploy, or browser execution.
  • A complete operating-system sandbox for live providers.

Documentation

Sign in through an official local CLI, run the preflight, then register a repository after reading the threat model.