Autonomous engineering agent
fractioncoder picks up tagged work items, writes the code, and opens the pull request. It keeps listening to comments, and it costs a fraction of a developer’s time.
Set a budget for every agent run.
How it works
Add one tag to a work item in Azure DevOps or a Jira issue. That is the whole handoff.
The agent reads the task, picks the repositories it touches and writes a plan; a second pass critiques it.
It implements the change and opens a draft pull request in each repository, linked to the task.
New comments on the task or the pull request wake it: it pushes a fix, replies, or asks a question.
fractioncoder can move the task through your own states as it works, and it stops when a task is closed.
Review comments on the pull request and notes on the task wake it with its full history intact.
Live logs, an interactive console, and VS Code in the browser. Stop or pause any run.
Create a workspace for each team or project. Every one has its own tracker, repositories, tools, secrets and agents.
Reporting
The Metrics page is built from the run records themselves, so completed work and spend always agree. See it as totals, per day, by outcome, and run by run. Admins see every workspace; everyone else sees the workspaces they belong to.
| Task | Title | Status | Spend | Duration | PR |
|---|---|---|---|---|---|
| AB#4821 | Fix retry in ingest job | done | $1.84 | 9m 42s | #214 |
| AB#4822 | Bump pyarrow, fix schema drift | done | $2.31 | 14m 05s | #215 |
| AB#4790 | Null keys in orders merge | monitoring | $3.07 | 31m 20s | #212 |
| AB#4783 | Add index to events table | done | $0.96 | 6m 11s | #209 |
| AB#4775 | Flaky date test in billing | done | $1.42 | 11m 48s | #216 |
Example data.
Cost
You configure the budget for each agent run. fractioncoder warns at 80 percent of the cap and pauses at it; a person decides whether to raise it. Spend and cost per run sit on the Metrics page.
Security first
Each workspace lists the tracker connection, repositories, tools and secrets its agents can use. An agent works with what you gave its workspace, and a repository allow-list keeps it to the repositories you name.
Agents get tokens limited to code, work items and wiki. No build, release or project administration. People sign in through OAuth 2.0 and OpenID Connect, so your identity provider stays in charge of who gets in.
Every change arrives as a draft pull request. A person reviews it and decides.
Nothing secret is baked into an image. Keys and tokens are injected when a container starts, and agents run as a non-root user.
Governance
Give each person a role in each workspace they work in. Someone with no role signs in and sees nothing. Admins manage workspaces, members and the settings that apply to all of them.
Your cloud, your keys
Self-hosted with Docker on Azure, AWS, GCP, or your own servers. You bring your own Anthropic API key, billed to you at API rates. Sign in through OAuth 2.0 and OpenID Connect, with Microsoft Entra ID or any other compatible provider. Secrets stay in your vault, never in images.