Issues in. Merged PRs out.

Orchestra runs AI agent pipelines on your GitHub repos. Plan, implement, review, and merge, each stage gated by verifiers.

Get Started See how a run works

Orchestra run detail for a completed Standard Change: the Orchestrator panel with every step done, a merged PR badge, and the run outcome with agents, models, sessions, and cost

One run, four gates.

Start a run from a GitHub issue. The Orchestrator writes a briefing for each stage, dispatches a specialist agent, and only advances when the stage's verifier passes.

Plan

An engineer agent reads the issue and opens a pull request carrying the implementation plan.

github_pr_opened

Implement

The agent pushes the change onto the plan's PR, behind the repo's build and test gates.

build_succeeds

Review

A reviewer on a different provider files findings. Rejected work loops straight back to implement.

github_review_approved

Merge

The Orchestrator merges the pull request once every gate has passed, and keeps the evidence.

github_pr_merged

Review findings loop the run back to implement until the reviewer approves. Every briefing, dispatch, and verdict lands in an append-only decision log.

A team of specialists, not a single bot.

Each stage is dispatched to a role-specific agent with its own briefing, skills, and model, running on Claude Code, Codex, or OpenCode.

Staff EngineerSoftware EngineerBackend EngineerFrontend EngineerDatabase EngineerDevOps EngineerSecurity EngineeriOS EngineerAndroid EngineerTech Lead

Code ReviewerQA EngineerQuality VerifierRelease ManagerTechnical WriterProduct ManagerProduct DesignerSpec WriterClassifierDecomposition

The harness around the run.

Pipelines your project owns

Standard Change, Quick, Advanced Feature, and Investigation ship built in. Clone one, then edit stages, roles, and gates per project.

Pipeline stage cards from a real run: Plan, Plan Review, Implement, and Review, each with its agent, model, time, cost, and a pass verdict

Verifier gates

A stage is complete when its verifier says so, not when the agent does.

github_pr_openedbuild_succeedstest_report_passinggithub_review_approvedforbidden_pattern_absentgithub_pr_merged

Components from the Hub

Install skills, agents, hooks, MCP servers, and rules per project. They materialize into the agent's worktree at launch.

Isolated sandboxes

Every session runs Claude Code, Codex, or OpenCode inside its own Docker container. Watch, stop, or restart the whole fleet.

Decisions on the record

An append-only decision log, live session activity, and per-run cost, kept for every run your team starts.

Put your backlog on the pipeline.

Connect a repo, pick a pipeline, and start your first run.

Get Started Sign In