Self-hosted control plane for AI agents. Definitions live in Git, every change becomes a PR, and runs, audit logs, and identity stay in your environment.

Tembo Agent Studio (TAS) treats agents like production software instead of editable prompts in a vendor console. Agent definitions live in a repository you own. New agents, edits, and "this run was wrong" feedback all produce the same reviewable artifact your team already trusts: a pull request.

• Git is the system of record. Agent definitions are files in your repo.
• Every change is a PR. Human or AI author, the output is always a diff.
• Runtime stays in your environment. Runs, audit history, and identity stay tied to your stack.
• Drift is governed. Agents can evolve without turning into opaque prompt sprawl.

Authoring is delegated to the Tembo Coding Agent Platform: TAS is the control plane, Tembo is the system that turns natural-language change requests into repository changes.

• Self-host the full stack with Docker Compose, from source or from published GHCR images.
• Sign in with Google, Microsoft Entra ID, or a generic OIDC provider.
• Connect a GitHub repository as the source of truth for agent definitions.
• Create, edit, and run agents in two frameworks: Pydantic AgentSpec and Cargo AI.
• Open PRs from chat-based authoring and from run feedback.
• Schedule agents with Automations or trigger them from external events.
• Manage per-user connections to external systems such as Slack, Gmail, Google Sheets, Notion, and GitHub.
• Inspect operational dashboards, run history, and append-only audit trails.
• Manage workspace membership with API-enforced roles.

See CHANGELOG.md for shipped work and ROADMAP.md for what is next.

cp .env.example .env

At minimum, set these required secrets:

• BETTER_AUTH_SECRET
• TAS_ENCRYPTION_KEY
• INTERNAL_API_TOKEN

Generate each with:

openssl rand -base64 32

You will also need at least one configured sign-in provider before anyone can log in:

• GOOGLE_CLIENT_ID and GOOGLE_CLIENT_SECRET, or
• MICROSOFT_CLIENT_ID, MICROSOFT_CLIENT_SECRET, and MICROSOFT_TENANT_ID, or
• OIDC_DISCOVERY_URL, OIDC_CLIENT_ID, and OIDC_CLIENT_SECRET

Set INSTANCE_ADMIN_EMAILS to bootstrap the first instance admin. Without it, the instance is invite-only and nobody gets the admin surface on first sign-in.

Build from source:

docker compose up --build

Or run published images:

docker compose -f compose.release.yaml pull
docker compose -f compose.release.yaml up -d

Once healthy:

• Web: http://localhost:3000
• API health: http://localhost:8080/health
• Postgres: localhost:5432

The API applies database migrations automatically on boot via sqlx::migrate!().

After the first admin signs in:

1. Create a workspace.
2. Connect the GitHub repository that stores your agent definitions.
3. Add an Anthropic and/or OpenAI API key in workspace settings so agents can run.
4. Optionally add a Tembo API key to enable chat-to-PR authoring and improvement requests.

The full zero-to-running checklist lives in Customer setup.

• Local or self-managed host: docker-compose.yml builds from source.
• Prebuilt images from GHCR: compose.release.yaml pulls ghcr.io/tembo/tas-api and ghcr.io/tembo/tas-web.
• Managed platforms: see the deployment guides for Railway, AWS, and Vercel.

Prerequisites:

• Docker or OrbStack
• Node 22+
• pnpm
• Rust 1.93+

Run everything with Docker:

docker compose up --build

Run only Postgres in Docker and develop the app code on the host:

docker compose up -d postgres

API:

cd api
cargo run

Web:

cd web
pnpm install
pnpm dev

Useful commands:

# web
cd web
pnpm lint
pnpm test

# api
cd api
cargo test

agent-studio/
├── web/        Next.js control plane UI
├── api/        Rust API, orchestration layer, and Postgres migrations
├── docs/       Astro Starlight user manual
├── agents/     Seed agent fixtures
├── context/    Phase docs, demos, and planning material
├── docker-compose.yml
└── compose.release.yaml

Repo-specific contributor guidance lives in AGENTS.md. If you change product behavior, update the user docs under docs/ in the same change.

The product manual is published at tembo.github.io/agent-studio and the source lives under docs/.

Recommended entry points:

• Introduction
• Getting started
• Customer setup
• Authoring agents
• Running agents
• Connections
• Automations & triggers

MIT