AgentOS: FastAPI for Agents

AgentOS turns your agents into a production API. One AI backend that serves every frontend.

1. Your product. Call the REST API from your app: run agents, stream responses, and manage sessions, memory, and knowledge.
2. AgentOS UI. Chat with agents, build new ones, and inspect sessions, traces, memory, and evals from the AgentOS UI at os.agno.com.
3. Coding agents. Manage the full agent development lifecycle (create, extend, improve, eval, review — and deploy) using the skills in .agents/skills/.
4. AI apps. MCP clients like Claude and ChatGPT can use your agents through the MCP server at /mcp.
5. Chat interfaces. Chat with your agents from Slack, WhatsApp, Telegram, and Discord.

Built on Agno. Everything runs in your cloud, your data lives in your database.

Get Started

Copy this prompt into your favorite coding agent. It sets up the platform and builds your first agent with you:

Help me set up my agent platform and build my first agent.

Clone https://github.com/agno-agi/agentos-fly into a folder called agent-platform, cd in, and run the setup-platform skill (in .agents/skills/).

Your coding agent drives the whole flow: it checks Docker, sets up .env, boots the platform, verifies the MCP endpoint, connects the AgentOS UI, and builds your first agent with you. Prefer to drive yourself? See Manual Setup.

Built for agents

This codebase comes with:

• Two platform agents that help you build and run the platform from your favorite AI apps like Claude and ChatGPT. Agent Builder creates agents, teams, and workflows using the AgentOS Studio. Platform Manager monitors and manages the platform: codebase questions, eval history, deployment checks, schedules.
• Coding-agent skills let Claude Code, Codex, Cursor, and other coding agents build, test, and improve the platform automatically — see Using the platform.

Trace data, agent code, evals, and system logs are all available to coding agents, so the platform can inspect and improve itself end to end.

Manual Setup

Step 1: Run locally

Prerequisite: Docker installed and running.

git clone https://github.com/agno-agi/agentos-fly agentos
cd agentos

# Configure credentials
cp example.env .env
# Open .env and set OPENAI_API_KEY

# Run the platform on docker
docker compose up -d --build

Confirm your AgentOS is running at http://localhost:8000/docs.

Step 2: Connect the AgentOS UI

1. Open os.agno.com and sign in.
2. Click Connect OS, enter http://localhost:8000 as the URL, name it Local AgentOS, and connect.

Step 3: Build your first agent

1. Click Chat under the Agent Builder agent and try the first prompt: "Build an agent that tracks AI news and writes a daily brief". Go through the agent development process.
2. Once created, click the Refresh button on the top right. You should now see the "Daily AI News Brief" agent in the Agents dropdown. Click the newly created agent.
3. Ask: "What's new with Anthropic?"

Step 4: Check platform health

Click Chat under Platform Manager and ask: "How healthy is the platform?" It answers from the codebase and runtime data — eval history, deployment checks, schedules, and the component you just built.

Run in production

You can run the platform anywhere that supports containerized images. This codebase comes with scripts to deploy the platform to Fly.io — and a coding-agent skill, /deploy-platform, that drives them for you and verifies the live platform at the end.

Prerequisite: flyctl installed and fly auth login completed.

1. Set up your production env

Create a new .env.production file for production credentials.

cp .env .env.production          # or cp example.env .env.production
# Edit .env.production with production values

Keeping a separate .env.production lets us use different values for local and production: different OpenAI keys, production-only credentials, a different Slack workspace.

2. Deploy

./scripts/fly/up.sh

This provisions the app and an unmanaged Fly Postgres on the same private network, pushes your credentials as Fly secrets, and deploys a single always-on machine. Deploys use fly deploy --ha=false on purpose: the Fly default creates two machines, which doubles cost and runs two in-process schedulers double-firing every cron. The script pauses and asks for a JWT verification key for authentication (see next section).

Cost note. Default sizing is shared-cpu-2x with 4 GB ($21/mo) plus a small Postgres machine ($4/mo). performance-2x (~$62/mo) is the dedicated-CPU option — edit fly.toml.

pgvector. Fly's stock postgres-flex image does not ship pgvector: sessions and memory work out of the box, but knowledge bases (RAG) need the extension. Set FLY_PG_IMAGE to a postgres-flex derivative with pgvector installed before running up.sh — the image is a two-line Dockerfile (FROM flyio/postgres-flex:17 + apt-get install -y postgresql-17-pgvector). Without it, up.sh prints a warning and everything except knowledge bases works.

3. Production Auth

Token-Based Authorization is on by default. Without a JWT_VERIFICATION_KEY or JWT_JWKS_FILE, the app refuses to serve traffic in production. The platform's job is to keep your data private, so the safe default is "refuse to start" without an authentication token.

Token-Based Auth gives you three things:

1. No public access. The server rejects requests without a valid token.
2. Per-request identity. Middleware parses the token and extracts the user_id, session_id, and custom claims. Each request is tied to a user and session, giving you auditability and traceability.
3. Granular permissions. User tokens can run an agent and view their own sessions. Admin tokens read everyone's sessions and test any agent.

During ./scripts/fly/up.sh, the app URL (https://<app>.fly.dev) is known before the first deploy, and the script pauses so you can mint the key before the app starts.

1. Open os.agno.com, click Connect OS → Live, and enter your Fly URL.
2. Name it Live AgentOS, flip Token-Based Authorization (JWT) on — the toggle is right on the connect panel — and connect. The UI generates your public key. (Already connected without it? Settings → OS & Security → Token-Based Authorization (JWT).)
3. Copy the public key.
4. Paste the full public key into the up.sh prompt. The script saves it into your env file for future syncs:

JWT_VERIFICATION_KEY="-----BEGIN PUBLIC KEY-----
MIIBIjANBgkq...
-----END PUBLIC KEY-----"

Heads up. Live AgentOS Connections are a paid feature. Use PLATFORM30 to get 1 month off. We are working on a free trial so you don't have to pay to try.

If you run non-interactively or skip the prompt, you can sync environment variables later with ./scripts/fly/env-sync.sh.

4. Register your production AgentOS to MCP clients

Re-run uvx agno connect, this time pointed at your deployed domain, to connect Claude Code, Claude Desktop, Codex, and Cursor to your production platform:

uvx agno connect --url https://<your-app>.fly.dev

For claude.ai and ChatGPT (web): add https://<your-app>.fly.dev/mcp as a custom connector in the chat app's connector settings. Leave the form's optional OAuth fields (client ID / client secret) empty. Click Connect and, on the consent page, enter the MCP_CONNECT_SECRET that up.sh generated during deploy (saved in .env.production).

5. Verify

You can check the logs on the Fly dashboard, or by running the following command (the app name comes from fly.toml):

fly logs

6. Redeploy after code changes

For updates from your machine, run the following command:

./scripts/fly/redeploy.sh

7. Sync environment variables

To re-sync environment variables, run the following command:

./scripts/fly/env-sync.sh

It reads .env.production by default (pass another file as an argument, e.g. .env) and pushes every variable as Fly secrets in one call — a single restart, no matter how many variables changed.

Tear down

./scripts/fly/down.sh

Destroys the app and its Postgres — including all data in the database.

Opting out of JWT (not recommended)

Set authorization=False in app/main.py and redeploy. Use this only inside a private network behind another auth layer. Without it, anyone who guesses your Fly URL can access your platform.

Using the platform

This platform is designed so that coding agents can drive the entire create → improve → evaluate → maintain lifecycle for you.

Create

Open your coding agent of choice (Claude Code, Codex, Cursor) and run:

/create-agent

It asks a few questions, generates the agent file in agents/, registers it in app/main.py, adds its description and quick prompts to app/config.yaml, restarts the container, and smoke-tests it live.

Improve

Improve your agents by running the following skills:

• /extend-agent — Add a tool, add a capability, refine the instructions, fix a known bug.
• /improve-agent — Claude simulates scenarios from the agent's INSTRUCTIONS, runs them against the live container, judges the responses, and edits until they pass.

Evaluate

Run the eval suite to check for regressions. The evals live in evals/cases.py, and run history shows up at os.agno.com next to your sessions and traces.

The evals run on the host machine, so set up the venv with ./scripts/venv_setup.sh && source .venv/bin/activate, then:

python -m evals --tag smoke      # fast checks of the self-driving surfaces
python -m evals --tag release    # broader pre-release confidence
python -m evals --name <case>    # one case while iterating
python -m evals -v               # stream the full run with rich panels

If a case fails, run /eval-and-improve — it diagnoses each failure, fixes what's in scope, and loops until green. And when you build an agent of your own, /create-evals writes its coverage: it mines your real sessions for scenarios and adds cases the scheduled eval run watches from then on.

Maintain

Because the repo is managed by coding agents, it moves fast. Run /review-and-improve before a release or after a refactor: it sweeps for drift between docs, code, and config, auto-fixes mechanical drift like stale paths and missing env vars, and flags anything bigger.

Connect more frontends (optional)

AgentOS comes with an MCP server at /mcp (enabled by setting mcp_server=True in app/main.py), so any MCP client can call your agents, teams, and workflows through tools like run_agent, run_team, and run_workflow.

Register your AgentOS with the MCP clients on your machine:

uvx agno connect

It auto-detects Claude Code, Claude Desktop, Codex, and Cursor and registers http://localhost:8000/mcp. After a successful connection, open one of these apps and ask:

can you access my agentos mcp?

claude.ai and ChatGPT (web). Hosted AI apps reach your platform over the internet and need an OAuth login. Deploy to production (above), add https://<domain>/mcp as a remote connector, and approve the consent page with your connect secret.

Environment variables

Variable	Required	Default	Description
OPENAI_API_KEY	yes	none	OpenAI key for models and embeddings.
RUNTIME_ENV	no	prd	dev disables JWT. Compose sets this to dev for local — never put dev in an env file that env-sync.sh pushes to Fly, or production serves unauthenticated.
JWT_VERIFICATION_KEY	prd	none	Public key from os.agno.com. Required when RUNTIME_ENV=prd, unless JWT_JWKS_FILE is set.
JWT_JWKS_FILE	prd	none	Path to a JWKS file; alternative to JWT_VERIFICATION_KEY for production JWT verification.
AGENTOS_URL	no	http://127.0.0.1:8000	Scheduler base URL. scripts/fly/up.sh sets it to https://<app>.fly.dev before the first deploy; set by hand only for a custom domain — re-running up.sh resets it to the generated fly.dev URL, so re-pin the domain (or re-run env-sync.sh) afterwards. Also the public origin OAuth metadata derives from when MCP_CONNECT_SECRET is set.
MCP_CONNECT_SECRET	no	none	If set (≥16 chars, e.g. openssl rand -base64 32), /mcp becomes its own OAuth 2.1 authorization server so claude.ai and ChatGPT (web) can connect; connecting asks for this secret on a consent page. Requires AGENTOS_URL. scripts/fly/up.sh auto-generates it on deploy. PAT and JWT bearers keep working alongside.
AGENTOS_MCP_SIGNING_KEY	no	none	Optional high-entropy signing-key material (≥32 chars) for OAuth tokens. Unset, a strong key is generated and persisted in the database. Rotating it invalidates outstanding tokens.
ENABLE_DEPLOY_CHECK	no	True	The reference deployment-check cron runs daily by default. Set False to disable; the workflow is runnable on demand regardless.
EVALS_TAG	no	smoke	Eval tag run by the run-evals workflow.
EVALS_CASE_TIMEOUT_SECONDS	no	90	Default per-case timeout for run-evals runs; applies only to cases that don't set their own timeout_seconds.
EVALS_SUITE_TIMEOUT_SECONDS	no	900	Whole-suite timeout for run-evals runs; per-case timeouts are the granular limit. The default bounds the smoke tag's worst case (incl. builder-case teardown).
PARALLEL_API_KEY	no	none	Authenticates the WebSearch Agent's Parallel SDK / MCP connection.
SLACK_BOT_TOKEN / SLACK_SIGNING_SECRET	no	none	Both must be set to enable the Slack interface.
DB_HOST / DB_PORT / DB_USER / DB_PASS / DB_DATABASE	no	matches compose	Postgres connection.
DB_DRIVER	no	postgresql+psycopg	SQLAlchemy driver.
AGNO_DEBUG	no	False	If True, Agno emits verbose debug logs. Compose sets this for dev.
WAIT_FOR_DB	no	False	If True, the entrypoint blocks on the DB before starting. Compose sets this.

Learn more

• Agno documentation
• AgentOS introduction
• Agno on GitHub. Drop a star if this is useful.