Devin got a new rival and its whole pitch is "same idea, but it lives in your AWS account."
lowkey that second half is the actual story.
OpenComputer landed on Hacker News this week as a Show HN and pulled 107 points and 32 comments. It's an Apache 2.0 cloud coding agent for teams. You mention @Computer in a Slack thread, it works on its own machine, your team can watch and steer it as it goes, and the result lands as a pull request. You can run all of it in your own cloud account (AWS, GCP or Azure, per the founders) or start on their hosted cloud.
my read: calling it "an open source Devin" undersells it. The agent isn't really theirs. The box is.
ok so what did they actually ship
The whole stack runs on a single VM, which they recommend locking down to outbound-only network access. On it sits a gateway called Calvin, a thin bridge that listens for Slack events and starts a session with Claude Code, Codex or OpenCode. Each session is isolated with rootless Podman. State goes to the VM's local disk first then to a git-backed store, so another VM can pick up the same session later. The rest runs on Cloudflare Workers and Durable Objects.
The architecture docs make one choice that's kinda spicy. Most platforms give every session its own sandbox that sleeps and wakes between turns. OpenComputer tried that and dropped it: too slow, cost grows with every session, and most sessions don't need one. A shared VM can handle hundreds of concurrent sessions in many cases, they say, and a session only gets a sandbox when it runs a heavy build or a preview environment.
On HN a founder added a cost claim: past roughly 640 session-hours a month, one always-on 4 vCPU VM is cheaper than per-session E2B sandboxes. Their math, not something I could check.
The hosted version has a 3-day free trial with 600 work minutes, and you can bring your own ChatGPT or Codex subscription.
the benchmark slaps, and it's also their benchmark
The headline is "3.3x faster than Devin." Here's where that comes from. The team gave each agent the same Slack prompt: bump zod to the latest version in four repos, run the checks and open a PR in each. They took the median over 5 rounds, and the test repo (speedbench) is public on GitHub.
| Agent | First push | All 4 PRs open, CI green | Runs completed |
|---|
| OpenComputer | 24s | 56s | 5/6 |
| Capy | 55s | 85s | 5/5 |
| Replicas | 53s | 87s | 4/9 |
| Hoplite | 69s | 95s | 4/6 |
| Devin | 72s | 185s | 5/5 |
tbh the gap looks real for this task. But a dependency bump is mostly "clone fast, install fast, push fast", exactly what a warm VM with fast disks wins at. It says way less about code quality, and since OpenComputer runs Claude Code or Codex under the hood, quality is mostly whatever you bring.
Also peep the last column. OpenComputer finished 5 of 6 runs, Devin 5 of 5. Not a scandal. Just a reminder that a vendor running its own race is gonna pick one it wins.
hot take though: speed is the right thing to race on. Their blog argues a slow cloud agent pushes you into fire-and-forget mode, you context switch, and your feedback lands after most of the work is wasted. imo that's a big reason people still run Claude Code on their laptop, and it's a sharper insight than the Devin comparison.
the HN thread was half feedback, half Devin drama
Plenty of commenters didn't know what Devin is. Rough for a headline built on it.
Then it got spicy. A commenter claimed Devin charges a premium markup and switched to a cheap default model. Commenters who said they work at Cognition pushed back: extra usage is billed at API pricing, Devin got 30 to 40% cheaper in Fusion and Normal mode in September, and the normal router uses Opus 5.5. Fun to watch, zero of it is about OpenComputer's code.
The useful bits were smaller. Several people said the name tells you nothing and is already taken, and the team said they'd take a look. Asked about OpenHands, the founders said OpenHands is an agent with its own harness while they're unopinionated on the harness and focused on the infra it runs in. Which is basically my take, confirmed by the people who built it lol.
Also no Discord support yet (it's on the roadmap), and "self-hosting" means your team's cloud account, not a home box running Ollama.
the question that didn't get answered
One commenter asked what the agent does without any check, and how you control it. In the thread as we fetched it, nobody replied to that one.
that's the part I'd care about if I was putting this on a team. A coding agent on a shared VM that anyone in a channel can ping gets interesting the moment it needs more than GitHub. Say you want @Computer to read a Sentry error, find the matching Linear ticket, then patch the code. Now there's a Sentry token and a Linear key sitting somewhere on that box.
This is where we'd reach for MewCP. Claude Code and Codex both connect to it with one config entry, the app keys stay in MewCP's vault and get injected per call, so the agent only ever holds a MewCP key. You can hand that agent a toolset that reads Sentry and files Linear issues but can't delete anything, and calls outside it get stopped before they reach the app. The use cases page has a Sentry-to-Linear triage flow. Big caveat: I haven't confirmed how OpenComputer passes MCP config into its sessions, so check that first.
stuff we still don't know
Whether 3.3x holds on real feature work, not a dependency bump. How permission checks work inside a session. What self-hosting really costs beyond the one VM plus Cloudflare Workers the founders describe. And whether the name survives the month.
where all this came from