Management layers on top of coding work. Agents (Claude Code, Codex, GitHub Copilot, Gemini) write the code; this repository is the Studio: a task board, agent pipelines, and project wikis that assign, gate, review, and account for it.
- Batch remote and local coding-agent work with task management in one place.
- Reduces cognitive load: task state, progress and evidence are visible instead of remembered.
- A management layer over autonomous agent runs: assign, gate, review, account.
- Works with your existing CLI subscriptions, or bring your own API keys.
- Runs fully local or distributed, your choice per project. Project chat follows the project's execution runner and shows its host, repository checkout, branch, and revision.
- Keeps the Task Server as the durable control plane while separately fenced Remote Review Executors inspect immutable result revisions.
- Runs review decisions, council reactions, post-processing, gate dispatch, and completion judging in the separate API-only Orchestrator Engine. Flow definitions and in-flight runs remain durable Task Server data, so restarting the Engine does not orphan work.
curl -fLO https://github.com/agent-orc/agent-studio/releases/latest/download/agent-studio-setup
curl -fLO https://github.com/agent-orc/agent-studio/releases/latest/download/SHA256SUMS
grep ' agent-studio-setup$' SHA256SUMS | sha256sum -c -
chmod +x agent-studio-setup
./agent-studio-setupThe default path needs Docker Engine on Linux or Docker Desktop with WSL2 on
Windows; --target native installs services instead on hosts without Docker.
The installer uses the version of the downloaded release, verifies its Compose
bundle, starts the full one-box stack, and opens the browser. Windows users
download agent-studio-setup.exe and verify it against SHA256SUMS.
See the install guide for Windows commands,
the native and connector profiles, unattended answers, update, rollback, uninstall, and remote topologies.
The one-box stack starts the Task Server, Engine, Studio BFF, compatibility API, web UI, and coding and review hosts. It creates restricted principal credentials in a named volume on first start and reuses them on later starts. Coding tasks additionally need provider and Git credentials mounted into the agent host. The default UI listens only on loopback; LAN exposure uses the optional HTTPS edge and an exact origin allowlist.
To run the same one-box stack from a source checkout (interim source-built path), bootstrap once and start the dev profile:
scripts/compose-distributed-bootstrap.sh
docker compose --profile dev up --build --wait task-server-dev orchestrator-engine-dev studio-bff-dev orchestrator-api-dev web-dev agent-host-distributed-dev agent-host-review-distributed-devSee the Docker operations guide for
updates, backups, product-managed credential rotation, credential mounts, and
the current route coverage limit beyond /api/v1.
scripts/scenario.sh --target inproc --level smoke runs the deployment
regression scenario: one seeded fixture driven end to end through bootstrap,
claim, run, and auto-review against the Task Server / Runner topology, no
Docker required. See
docs/operations/testing/deployment-scenario.md.
Agent Studio is part of the agent-orc ecosystem. It uses Chat for coding-agent conversations and sits alongside Runner for hardened CLI execution, Token Economy for model pricing and usage accounting, and Quality Studio for layered code review.
For future product direction, see the roadmap. For architecture, ADRs, contracts, and the full documentation index, see agent-orchestrator.dev.
Licensed under the Apache License 2.0.