OpenComputer Open-Sources Its Devin Alternative: One Durable VM, First Push in 24s
On this page 7 sections
TL;DR
OpenComputer, an open-source alternative to Devin, took its Show HN turn on Nov 20 with the founders answering questions in the thread. Mention @Computer in a Slack thread and it works on its own computer, running your choice of Claude Code, Codex, or OpenCode, and the result lands as a pull request. The stack is Apache 2.0 from the compute layer up, and you can run it in your own AWS, GCP, or Azure account.
In the team's published benchmark, first push arrives in 24 seconds and all four benchmark PRs go green in 56, against 185 for Devin. The speed traces to one design bet: a durable VM per team, with sessions isolated inside it, instead of a fresh sandbox for every task.
What shipped
The platform itself is about two months old: an agent cloud in public beta since late September, the @Computer Slack agent since an October 30 post, and the open-source announcement on November 13. Igor Zalutski, Mohamed Habib, and Utpal Nadiger signed both posts, and two of them worked the HN thread, which sat at 69 points and 21 comments within hours. The repo counts 746 stars, 80 forks, and 95 open issues at time of writing, so calibrate "Devin alternative" accordingly.
Two ways to run it. On OpenComputer Cloud you get a 3-day trial with 600 work minutes, and you can bring the ChatGPT or Codex subscription you already pay for. Self-hosting is one VM in your cloud account plus the open web stack, with guides for AWS, GCP, and Azure. The founders were blunt in the thread that this means a team cloud account for now; a homelab box running Ollama is on the someday list.
The benchmark, and its fine print
The task is deliberately dull: bump the zod dependency to the latest version in four small repos, run the checks, open a PR in each. The repos and setup are public. Every provider got the same Slack prompt, five rounds, with the clock running from the message timestamp to the PR appearing on GitHub. Dull is the point here: a trivial task separates infrastructure latency from model ability.
Read the numbers with the caveats attached. OpenComputer wrote the task, ran the clock, and published the results, so treat it as a reproducible claim rather than a neutral verdict. The blog discloses that several rivals needed a nudge between rounds because a closed PR from the previous round made them refuse the job, which the authors blame on memory interference. The runs-completed column deserves a glance too: Replicas finished 4 of 9 attempts, OpenComputer 5 of 6, everyone else 5 of 5.
Devin's 185 seconds for four green PRs is the eye-catcher, and still faster than most standups. The first-push column matters more day to day, because that delay decides whether you keep steering the agent or drift into another tab while it works. That is the whole argument of the benchmark post: cloud agents became "background agents" by default, and the authors want the interactive loop back.
Why it's fast: one warm VM
Most cloud coding agents boot an ephemeral sandbox for each task or session. OpenComputer runs the agent harness and the sandbox together on a single durable VM, one fleet per organization, sized from one machine up. A per-session sandbox is hot-desking: every task starts by unpacking your tools and waiting for the machine to warm. The durable VM is your own desk, chair still warm, repo already checked out.
The path a task takes is short. Slack events land in Cloudflare Workers and Durable Objects; your VM never accepts inbound connections, it dials out over a websocket, so your security team gets an outbound-only machine they can lock down. On the VM, a gateway called Calvin spawns one rootless Podman container per Slack thread, with the checkout held in a git worktree. Session state is written to local NVMe first and committed to a git-backed store after every turn, so a dead VM costs you a resume, not a restart.
Asked how this differs from OpenHands, founder Utpal Nadiger kept the boundary crisp: "OpenHands is an agent with its own harness. We're unopinionated on the harness and more focused on the infrastructure it runs in."
That shows up in the local test too. Side by side against Claude Code on a MacBook Pro M2, the cloud session matched or occasionally beat the laptop, which the team credits to sitting near the inference provider. Whether you see the same depends on your own ping to the region.
The economics of warm
The founder post includes a cost crossover: past roughly 640 session-hours a month, one always-on 4 vCPU VM runs cheaper than paying for an E2B-style sandbox per session. That is vendor math with vendor pricing behind it, so rerun it for your workload before you migrate anything. The shape of the argument holds regardless: a warm machine makes startup free, and idle VM minutes cost less than provisioning a sandbox per PR.
Model spend stays where it already lives. You bring the harness, so the subscription you already pay does the thinking, and the cloud side meters in work minutes.
Reasons to hold off
- The benchmark is vendor-run on a task the vendor wrote. The harness is public; rerun it before you quote the numbers.
- The code has been public for a week and carries 95 open issues.
- Self-hosting means a cloud account, not a spare box under your desk; no Ollama path yet, and Discord is only on the roadmap.
- Slack is the primary surface, with a web UI and CLI as fallbacks. Teams living elsewhere should wait.
Key Takeaways
- OpenComputer is an Apache 2.0, self-hostable alternative to Devin: mention @Computer in Slack, get a PR, with Claude Code, Codex, or OpenCode doing the work underneath.
- Its published benchmark shows 24s to first push and 56s to four green PRs, versus Devin's 72s and 185s; the speedbench harness is public, so verify before repeating.
- The design bet is one durable VM per team, rootless Podman isolation per Slack thread, git-backed session state after every turn, and outbound-only networking.
- Founders peg the cost crossover at about 640 session-hours a month, where one 4 vCPU VM beats per-session sandboxes.
- You can bring your existing ChatGPT or Codex subscription; the cloud trial is 3 days and 600 work minutes.
- Vendor benchmark, week-old code, Slack-first: strong infrastructure ideas, unproven at your scale.
Sources: OpenComputer on GitHub, Show HN thread, OpenComputer benchmark post, open-source announcement, self-hosting docs, speedbench repo