Open source alternative to Devin
Devin runs on Cognition's infrastructure, takes a ticket, and works until it has a pull request. When it lands the task it is impressive. The reasons to look elsewhere are structural: your code runs on someone else's machines, the unit of billing is compute rather than outcome, and reviewing what came back is still entirely your problem.
Closed source, hosted only. There is no self-hosted build.
Web, with Slack and GitHub integrations.
Priced by agent compute unit, on top of a monthly plan.
Why people replace Devin
Your source runs on their infrastructure
The whole model is remote execution. For code that cannot leave your network, there is no configuration that fixes this.
You pay for attempts, not results
Compute units are consumed whether the run lands the task or wanders. Budgeting against that is hard.
The pull request still has to be reviewed
A long autonomous run produces a large diff. Autonomy moved the writing; it did not move the reading.
One vendor, one model
You get Cognition's agent and Cognition's choices. If a different model is better at your codebase this month, that is not a lever you have.
The layer above your agents
The agent runs where you say
Your laptop, your server, your air-gapped box. Aura drives agent CLIs on machines you control, and the engine is local by default.
A verdict on the pull request
Every session carries the goal it was given. At the end Aura reports which parts of that goal are wired, which are partly wired, and which are missing, before anyone opens the diff.
Your choice of agent, and the same record
Claude Code today, Codex tomorrow, a local model next quarter. The semantic history does not care which one wrote the change.
You pay the model provider directly
Aura is free and Apache 2.0. There is no second meter between you and the tokens, and the per-response cost is shown as it accrues.
Aura is not better at everything
- Devin needs no machine of yours. With Aura, the compute is yours to provide, which is either the point or a cost depending on your setup.
- Its ticket-to-PR flow is more automatic out of the box.
- Devin will keep grinding on a task for hours with nobody watching. Aura can do that through its crew runner, but it takes more setting up.
The other open-source options
If Aura is the wrong shape for what you need from Devin, these are the projects worth reading first.
How the move actually goes
- 01Pick the machine the agent should run on: a laptop, a spare box, or a cloud VM you own.
- 02Install Aura and the agent CLI you want to drive.
- 03Turn each ticket into a goal sentence. This is the same input Devin takes, written down where the proof can use it.
- 04Run the session, then read the semantic diff and the goal verdict rather than the raw pull request.
- 05For unattended work, use the crew runner: a task graph, one worktree per agent, and a proof at each completion.
Aura and Devin
Can Aura work autonomously like Devin?
Yes, through the crew runner: tasks are laid out as a dependency graph, agents claim what is unblocked, and each completion is tied to the goal it delivered. The difference is that the machine is yours and each finished task carries a proof.
Does anything leave my network?
Only the model calls, which go to whichever provider you configured. The engine, the index and the history are local.
What does it cost?
The software is free and Apache 2.0. You pay your model provider. Aura meters the spend per response so the number is not a surprise at the end of the month.
How do I know the agent actually finished the task?
The proof. Aura checks the goal against the AST graph of what changed and reports wired, partly wired, or missing, with the gaps named.