OpenComputer, an open-source, self-hostable alternative to Devin, landed on Hacker News this week as a Show HN from one of its founders S4. The "Devin alternative" framing is the least interesting part. What deserves attention is the deployment posture: where the agent runs, who can watch it, and what it is allowed to touch.
The posture is the product
Per the project's README, you mention @Computer in a Slack thread with a task, it works on its own computer while your team watches and steers it, and the result lands as a pull request S1. Self-hosting is deliberately boring: one VM locked down to outbound-only network access, a thin gateway named Calvin that starts a Claude Code, Codex or OpenCode session per thread, each session isolated in rootless Podman, state written to the VM's disk and then persisted to a git-backed store, and a control plane on Cloudflare Workers and Durable Objects S1.
Read through a forward-deployed engineering lens, this is the point: the worker lives inside the customer's channel and cloud account rather than behind a vendor portal. The seams are clean. The gateway is where you swap harnesses, Podman is where sessions die, git is where state survives. That is the same discipline teams already apply to any service that touches production code.
Latency is a delivery spec
The team's own benchmark post is blunt: "We lose interactivity with background agents as of today!" S2. Slow replies push engineers into context switches, and feedback that arrives only when a project is done wastes most of the work in flight S2. Their measured claim: a median 24 seconds from Slack message to first push and 56 seconds to four pull requests with CI green, against 72 and 185 seconds for Devin S2. Caveats matter: it is a vendor-run benchmark across five rounds of a single task type, so treat it as a claim to reproduce, not as proof S2.
The takeaway survives the caveat. A deployed agent needs an interaction budget, treated like an SLO. If answering a teammate costs more than the price of switching away to something else, the agent's throughput gains quietly evaporate.
Fit the model to the harness you ship
Hugging Face's harness-adapters post, published this week, shows the other half of the delivery stack. Qwen3.8-27B scored 39% on Terminal-Bench 2.1 under the pi harness at roughly 2.9K tokens per task; with the weights frozen and a custom RL-trained LoRA adapter added, it reached 55% pass@1 at about half the token usage S3. Training ran the harness exactly as it ships, and on the tasks that both the base model and a merged "super adapter" solve, with thinking off, the merged version cut tool calls by 53–73% and tokens by 44–60% S3. Notably, the specialist adapter still beat the merged one under pi on SWE-bench Pro, 64.0% versus 46.0% S3.
For delivery teams the lesson is concrete: when your gateway can swap harnesses, harness choice becomes a deployment decision, and the tuned model belongs to the harness actually running in production. Change the harness, and you re-tune or re-select the adapter rather than assuming benchmark gains transfer.
None of this requires exotic infrastructure. Isolate every session, restrict egress, keep state in git, budget response latency, match the model to the harness — and rerun the vendor's benchmark yourself before quoting it.
Sources and editorial method
This AI-assisted report synthesizes the sources below. Automated source checking is not independent human verification; follow the original reporting for context and updates.
- S1: GitHub - diggerhq/opencomputer: An open-source, faster Devin alternative
- S2: What happens when cloud agents are as fast as your local machine? – OpenComputer
- S3: Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost
- S4: Show HN: OpenComputer – an open source alternative to Devin
Common questions
What is OpenComputer?
An open-source, self-hostable coding agent: you tag it in a Slack thread, it works on its own computer, and the result lands as a pull request. Sessions run isolated in rootless Podman on a VM with outbound-only network access.
Why does agent response latency matter for delivery teams?
The OpenComputer team argues background agents lose interactivity rather than capability: delayed replies force context switches and late feedback wastes in-flight work. Treat response time as a delivery budget, like an SLO.


