Structured execution for AI-driven work.
Obra turns one objective into a structured plan, then orchestrates parallel agents that verify their own work before it converges into a finished outcome — a reviewed commit, or a delivered document, depending on the work. Runs on the providers and machine you already use.
Mission in. Work out.
Obra reads a single objective and runs the whole arc. The work gets decomposed into a plan, dispatched to agents that verify as they go, and converged into one reviewable result. That turns a goal into structured delivery instead of one long, unaccountable session.
Define the objective
Hand Obra one goal in plain language, scoped to a workspace. You stay the architect.
→Decompose into structure
The objective is broken into a nested plan: goals first, then the concrete tasks under each, so the work is legible up front.
→Run in parallel, verified
Agents implement, test, audit and review in parallel lanes, each owning one slice of the plan.
→Converge to a verified result
The parallel streams come back together as one reviewed result, versioned and traceable in your repository.
Watch it run, in plain language.
Obra narrates every step as it runs and records what changed, so you can follow the work and revert anything you didn't want. No black box.
Specialized agents, each accountable for one job.
Obra doesn't ask one model to do everything at once. Dispatches purpose-built agents and coordinates them: implementation, testing, security, and review run as distinct, verifiable roles.
Reads the diff and the surrounding code, then reports the issues worth acting on.
Scans changes for vulnerability classes — the risk surface a fast implementer is prone to miss.
Authors coverage for what changed, so a slice arrives with its own regression net.
The same engine runs two domains — software against your repository, and business pipelines for any knowledge work. Toggle the console to watch each one run.
Run it where you work — terminal, canvas, or API.
The same orchestration core is reachable from wherever you work: a CLI for the keyboard, a visual studio for the plan, and a programmable gateway for everything else.
The obra CLI
Your home base: start a session, inspect a run, and drive the whole platform from one command.
Workflow Studio
A visual web client for authoring workflows, monitoring runs, and editing the plan with AI.
Gateway API
A hosted HTTP and WebSocket gateway with an event stream and webhooks for integration.
MCP server
Expose Obra to any MCP-compatible client over stdio — Claude Desktop, editors and beyond.
Desktop app
A native macOS app with guided onboarding and gateway lifecycle management (distribution pending).
Multi-provider core
Drives Claude Code, OpenAI Codex, Gemini and Ollama — bring your own provider and your own auth.
Your code stays local. Only orchestration crosses the wire.
- ΩCode never leaves your machine. Agents run locally against your workspace; only orchestration metadata reaches the Obra service.
- ΩBring your own provider. Obra drives the provider CLIs you already authenticate, using your own keys and accounts.
- ΩCommits in your repository. Work converges into your own version control, with a history you can trace and revert.
- ΩVerification in the loop. Review, tests, and audits run as part of execution, gating the work before it lands.
Install with uv. Run your first objective.
Obra installs as a single uv tool. Authenticate once, point it at a provider you already use, and hand it a goal.
$ uv tool install obra # uv-only — the supported path $ obra login # accept terms + browser auth $ obra doctor # check provider CLIs + config $ obra run "Add OAuth and harden the session layer" Ω decompose · dispatch · verify · converge…
Obra drives a provider CLI you already run — Claude Code,
OpenAI Codex, Gemini or Ollama.
You bring the access, whether a cloud account or a local model, and obra setup picks which one Obra calls.
Design the mission. Let Obra orchestrate the delivery.
Obra means a finished work, and that's what you get.
Beta research preview, access-gated · bring your own provider · read the docs