How it works

From your task to your PR

You do three things — connect, hand off, review. The agent does the rest while you're away. Here's the full loop.

The FlareCode agent workspace: fleet rail, a live build session with reasoning, edits and tests, and the diff review panel
the agent's workspace — your fleet on the left, the build loop in the middle, the diff on the right

▸ You

  1. 01

    Connect a repo

    Sign in with GitHub, then install the FlareCode app on the specific repo you want the agent to work on — per repo, never across your whole org.

    It gets only the four permissions it needs to open a Pull Request: read your code, write to a branch, read the default branch, and read CI status. Nothing org-wide, no admin access.

    Revoke access any time from your GitHub settings; we delete the install record on our side within 60 seconds.

  2. 02

    Hand off the task

    Open the agent for that repo and describe the task in plain English — as specific as a bug report you'd file yourself.

    Nothing is bound to that browser tab. Close the laptop and walk away; the agent keeps working on the server while you do something else.

    ▸ The agent

    while you're away

    ↳ hand off

    1. step 1

      Plan

      Reads your code and breaks the task into steps.

    2. step 2

      Code

      Edits files in an isolated, encrypted sandbox.

    3. step 3

      Test

      Runs your tests and fixes what fails — no infinite loops.

    4. step 4

      Open PR

      Pushes a branch and opens a Pull Request for you.

    ↳ opens a Pull Request

    Every task runs in its own isolated sandbox with a per-task spend limit you set. Hit the limit and we stop the task and refund the platform fee — so the cost of every task stays predictable.

  3. 03

    Review & merge

    When the agent is done it opens a Pull Request on a branch named flarecode/task-<id>. It never pushes to your default branch.

    You get a notification — email, GitHub mobile, or Slack. The PR shows the task you wrote, what changed, and the test output.

    Approve and merge from anywhere, or close it if it's wrong. The decision is always yours.

The loop, surface by surface

What each step looks like

The composer with Plan, Build, and Auto modes and a model picker

1 · Hand off

Describe the outcome, pick a mode

Type the task in plain English. Choose Plan to review a plan first, Build to go straight to code, or Auto-pilot to let it run end-to-end — and pick the model per task. Or type /loop to hand off a standing objective and let a fleet keep shipping toward it, on a schedule or until it's done.
Tool-call cards: a test run, a self-check, and a file edit with a diff

2 · Build & verify

It edits, runs your tests, and checks itself

Watch the agent reason, edit files, run your test suite, and self-check the running app, every step in the timeline. A red test bounces the task back for another pass, not a green checkmark it didn't earn.
Side-by-side git diff with additions highlighted and an Open PR button

3 · Review

Read the diff, open the PR

See exactly what changed in a real side-by-side diff. When it looks right, open the Pull Request from the panel. It merges through your normal GitHub review, with you as the approver.
A rail listing several agents with live status dots and nested sessions

Across repos

Run a whole fleet at once

Point an agent at each repo and they work in parallel in the cloud. Switch between them in one tab and pick up any PR from your phone.

What persists

Your work is saved; the compute isn't

Your workspace is encrypted and continuously backed up — an idle pause, a restart, or a from-scratch project never loses work. We keep it only to make your work resumable, never to resell or train on it, and it's destroyed when you destroy the agent.

The compute that runs each task is ephemeral, torn down when the task ends. What remains: your saved workspace, your PR on GitHub, sanitized logs (30 days), and billing metadata (90 days). GitHub stays your source of truth. See the full security posture →

Get started

Try it

Connect a GitHub repo, hand off a task in plain English, and get a reviewed pull request.

Start free →

free to start · usage billed at cost · cancel anytime

See what it costs →