What is o8?

o8 is a local-first desktop app that runs a team of AI coding agents under one orchestrator, with a governance layer that reviews every change before it merges.

The short version

o8 is a control room for autonomous engineering. You describe what you want; one orchestrator decomposes the work, dispatches it to a fleet of coding agents, and a review gate sits between those agents and your main branch. Nothing merges unreviewed, and nothing risky merges without your sign-off. It runs on your own machine with the model subscriptions or keys you already use.

It is the layer abovethe coding tools, not another editor. If a developer asks “what runs a team of coding agents locally, with a review gate,” that is what o8 is built to be.

How it works

The orchestrator breaks a task into packets and dispatches each one to a worker agent running in its own isolated git worktree. The agents can’t reach each other and can’t touch main until the gate allows it. Run ten at once — when two reach for the same file, you see the collision before anyone ships.

• Orchestrator — decomposes your intent, dispatches the fleet, and reports back.
• Parallel fleet — every agent works in a sealed-off git worktree.
• Review gate — every diff queues for review; approval is a decision, not a default.
• Merge gate — rebase, typecheck, and head-SHA lock, then merge in order.
• Conflict detection — collisions between lanes surface before anyone ships.
• Engineering Brain — agents ask the repo instead of grepping; answers come with citations.
• Organizational memory — when the same fix lands three times, o8 proposes a rule you accept or reject.
• Audit trail — every packet, merge, and event is recorded and replayable.
• Mobile approval — read transcripts and approve or reject from your phone.
• Symon — a voice operator that dispatches through the same gate.

What makes it different

Tools like Cursor, Copilot, and Claude Code put one model inside your editor. o8 coordinates many agents and adds the part they leave out: a governance layer — approvals, audit, and organizational memory — so you can let agents run without losing oversight. The agents do the work; you run the room.

Which models it runs

The current public release has 16 dispatchable agent runtimesbehind one adapter contract. The orchestrator and worker are separate choices, so one runtime can hold the goal while a different runtime and model execute each packet. Workers use the subscriptions and keys you bring. Each installation’s o8 chat also includes a limited hosted inference allowance without sign-in, and you can switch back to your own key whenever you want.

Where it runs, and what it costs

o8 has a signed, notarized release for macOS 14+; the current build is Intel and runs on Apple Silicon via Rosetta. Windows 11 has public MSI and EXE previews with checksums and build attestations; publisher signing remains open. Linux has public AppImage, deb, and rpm previews; the AppImage is GPG-signed, while desktop verification and update delivery remain open. The app is open source and free forever with the model subscriptions and keys you already use, including the full app, governance gate, Engineering Brain, and canvas. Local repositories and installed or bring-your-own model tools do not require an o8 account. Each installation gets a limited hosted model allowance, provisioned on demand when first used. Sign-in is for portable account services such as sync, the web console, mobile relay, and Pro. Optional Pro pays for the managed layer and provides our most generous daily fair-use limits for life.

Your repositories, sessions, approvals, and memory stay on your machine in a local database (~/.o8). Requests use either the providers you configure or o8’s hosted inference when you choose that route. More in the docs, pricing, and FAQ.