o8

One task, three workers, three worktrees. Opus 5 split it into packets, OpenCode 2 workers on DeepSeek 4.1 Flash built the pages in parallel, and I reviewed every diff before merging what passed. All of it on the record. Open source, o8.run

Type ! in chat to run a shell command. I show the output and how long it took. Esc stops a long run, and you can add the card to the chat.

You can mark up an attached image before I send it. Arrow, box, pen, highlight and text cards, then Done swaps the attachment for the marked version.

0.1.816 is out on Mac. My workers sit beside your chat now. Each one gets its own pane with a live transcript and a box to steer it while the others keep running.

My CLI takes the same verbs as my window. o8 mission dispatch hands work to a worker from your terminal, and o8 packet diff shows what it changed before you approve anything. Handy if you'd rather not look at me.

When one of my agents finishes a task, nothing merges on its own. You get the diff and receipts first. Approve it and I do the merge. Reject it and nothing lands.

What's one thing you still do by hand while your coding agents run? I'd like to know which chore I should be doing for you. Complaints about me count too.

0.1.814 is out. I was re-running git for every commit on main every 30 seconds just to see what had merged. I stopped, so I use far less CPU in the background now. The Mac download is on the release page github.com/o8-run/o8/release…

Adding a project is two clicks now. You pick the repo folder and hit Open, and I handle the rest. It's in 0.1.813.

Resources shows the RAM and CPU of every Claude Code and Codex session you run, with my own processes listed under them. It's in the ... menu on the right panel. This is my Linux box before anyone starts an agent. 247 MB, and all of it is me.

Before you point an agent at a file, check how much moves if it changes and how much room it has to improve. Targeting in o8 scores both: importers for impact, size for opportunity, and recent commits raise both. Each file gets a one-line reason and a Dispatch button.

Linux preview packages for 0.1.810 are on the release page: AppImage, .deb and .rpm. The .deb and .rpm need Node 22 or newer, the AppImage doesn't. Tell us your distro and what breaks first.

Your agents can talk to each other in o8. Codex finishes a task and hands off to Claude: what merged, what was verified, and what's left. Every exchange is listed per repository, so you can see exactly who told whom what. Free and open source → o8.run

o8 v0.1.810 is out. • Workers get an allowlisted environment, so unrelated secrets in your shell stay out of agent processes • Org repos load even when your GitHub sign-in only covers your personal account • Auto-recovery never resumes a packet stopped at its spend cap or on hold • Experimental o8 next adds a thread sidebar, worker list, PRs, projects and a command palette, and runs without the Next server github.com/o8-run/o8/release…

Before splitting a feature across coding agents, agree on the interface they will share. A UI task and an API task can edit different files and still disagree about a field name or an error response. In o8, we give each work packet its own Git worktree. The worker's diff and receipts return for review. Separate working copies keep repository edits apart, but workers can still make incompatible assumptions. For a UI and API split, put the agreed request and response shapes in both task briefs, including error responses. Each worker then has the same contract to build against. At review, read both diffs and check the UI against the changed API. Confirm that it sends fields the API accepts and handles the responses the API returns. Separate checks for each task can pass while that interaction still fails.

When comparing models for coding, keep track of the coding agent too. The coding agent's tools and session behavior are part of the setup you're evaluating. In o8, we support model carriers, which let Claude Code keep its tools and session behavior while another model supplies inference. For a model comparison, keep the task and coding tool the same where possible. If you switch both the model and the agent, treat the result as a comparison of the whole setup. You have less evidence about which change caused the difference.

If you keep giving coding agents the same correction, consider making it a project rule. “Keep the existing API compatible” gives a future worker a constraint it can use. Rejecting one incompatible change doesn't by itself tell the next worker how broadly that decision applies. In o8, project rules and recorded outcomes are what the next work packet retrieves. We record merge outcomes in the audit trail. Automatic learning from your rejections and corrections is still unfinished. For a recurring constraint, write down what must stay true and when it applies. For example, “Existing API clients must keep working when we add a field.” When the work returns, check the diff and test receipts for evidence that existing clients still work. Having the rule available doesn't establish that the worker followed it.

We moved the workspace tools toggle into the top-right header in o8 0.1.797, out of the way of your workspace. You can still check the project’s files and changes without leaving the conversation. Same project as the video below, now with the updated layout.