AgentOS: FastAPI for Agents

July 27, 2026 · View on GitHub

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

Your product. Call the REST API from your app: run agents, stream responses, and manage sessions, memory, and knowledge.
AgentOS UI. Chat with agents, build new ones, inspect sessions, traces, memory, and evals from the AgentOS UI at os.agno.com.
Coding agents. Manage the full agent development lifecycle (create, extend, improve, eval, review) using the skills in .agents/skills/.
AI apps. Use your agents from Claude and ChatGPT using the MCP server at /mcp.
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-helm 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:

Chief, your team's mascot. "Chief, we're going with PlanetScale over RDS." "Chief, zak ran a good launch." Tell it anything — decisions, who's on what, what you learned — and it files the who and the why, learns how you work, and connects the dots when someone asks what's happening. Every frontend talks to the same Chief: what you tell it in Slack is there when you ask from claude.ai or ChatGPT.
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 understands, monitors, and explains 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-helm 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

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

Step 3: Build your first agent

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.
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.
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.

Step 5: Meet Chief

Click Chat under Chief and tell it what you're working on: "Hey Chief — I'm building a support bot." It files the who and what as entities, the why as notes, and what it learns about you stays yours. From then on, tag it in from anywhere — this UI, Slack, claude.ai, ChatGPT — and ask "Chief, what's happening?": same thread everywhere.

Run in production

You can run the platform anywhere that supports containerized images. This template deploys to any Kubernetes cluster — cloud-managed (EKS, GKE, AKS) or your own — with the Helm chart in charts/agentos and a single script; the /deploy-platform coding-agent skill guides you through it.

Prerequisites: kubectl pointed at your cluster, Helm 3+, a container registry the cluster can pull from, and OpenSSL.

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. The image

The chart defaults to the official agnohq/agentos image — the reference platform exactly as in this repo (latest, plus agno-<pin> tags for exact runtimes). The moment you customize anything — a new agent, edited instructions — build and push your own and point the chart at it:

docker build -t <registry>/agentos:v1 .
docker push <registry>/agentos:v1

(Testing on a local kind cluster instead? docker build -t agentos:kind . && kind load docker-image agentos:kind — see Local dry run on kind.)

3. Deploy

./scripts/k8s/up.sh                                              # official image
IMAGE_REPOSITORY=<registry>/agentos IMAGE_TAG=v1 ./scripts/k8s/up.sh   # your own build

This helm-installs the chart into the agentos namespace of your current kubectl context (the script shows the context and asks first): the API deployment — one replica by design, the in-process scheduler must not run twice — plus in-cluster Postgres with pgvector and its volume. The script pauses and asks for a JWT verification key for authentication (see next section). To publish it behind your ingress controller, add INGRESS_HOST=os.example.com (and optionally INGRESS_CLASS=nginx); AGENTOS_URL then points at that host, otherwise the scheduler uses the in-cluster service DNS, which works out of the box.

Bringing your own Postgres instead? It must have the pgvector extension available — install with postgres.enabled=false and the externalDatabase.* values (see charts/agentos/values.yaml).

4. 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:

No public access. The server rejects requests without a valid token.
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.
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/k8s/up.sh, the script pauses so you can mint the key before the app starts.

Open os.agno.com, click Connect OS → Live, and enter your AgentOS URL (your ingress host — or a tunnel while testing).
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).)
Copy the public key.
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/k8s/env-sync.sh.

5. 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-agentos-domain>

For claude.ai and ChatGPT (web): add https://<your-agentos-domain>/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; deployed without INGRESS_HOST? set MCP_CONNECT_SECRET and a public AGENTOS_URL in .env.production and run ./scripts/k8s/env-sync.sh).

6. Verify

kubectl rollout status deployment/agentos -n agentos
kubectl logs deploy/agentos -n agentos -f

No ingress yet? Port-forward and open http://localhost:8000/docs:

kubectl port-forward svc/agentos 8000:8000 -n agentos

7. Redeploy after code changes

Build and push a new tag, then roll the release to it:

docker build -t <registry>/agentos:v2 . && docker push <registry>/agentos:v2
IMAGE_TAG=v2 ./scripts/k8s/redeploy.sh

Immutable tags keep rollbacks one helm rollback away. Re-pushing the same tag and running ./scripts/k8s/redeploy.sh without IMAGE_TAG restarts the pods instead — that only picks up the new image if it actually reached the cluster.

8. Sync environment variables

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

./scripts/k8s/env-sync.sh

Changed secrets roll the pod automatically (the deployment carries a secret-checksum annotation). The script syncs the connection and secret keys (including JWT_JWKS_FILE); other knobs (ENABLE_DEPLOY_CHECK, EVALS_*) are chart values — set them via extraEnv and helm upgrade.

9. Tear down

./scripts/k8s/down.sh

Uninstalls the release and deletes the database volume, including all data. The namespace is kept — it may be shared; the script prints the optional delete command.

Local dry run on kind

The whole template runs on a laptop-local kind cluster — the same flow the family E2E uses:

kind create cluster --name agentos
docker build -t agentos:kind . && kind load docker-image agentos:kind --name agentos
printf 'RUNTIME_ENV=dev\n' > .env.production && grep '^OPENAI_API_KEY=' .env >> .env.production
IMAGE_REPOSITORY=agentos IMAGE_TAG=kind IMAGE_PULL_POLICY=Never ./scripts/k8s/up.sh
kubectl port-forward svc/agentos 8000:8000 -n agentos   # then open http://localhost:8000/docs
./scripts/k8s/down.sh --yes && kind delete cluster --name agentos && rm .env.production

RUNTIME_ENV=dev disables JWT so nothing needs minting — never sync a dev env file to a real cluster.

Opting out of JWT (not recommended)

Set authorization=False in app/main.py and redeploy. Use this only inside a private VPC behind another auth layer. Without it, anyone who reaches your AgentOS 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 and its real usage recorded in the database, 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.

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 a real cluster, 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. The chart resolves it automatically (explicit value > ingress URL > in-cluster service DNS); set by hand only for a custom domain or tunnel. 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 a public `AGENTOS_URL`. `scripts/k8s/up.sh` auto-generates it when the deploy has a public URL (`INGRESS_HOST` or `AGENTOS_URL`). 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 Chief's and the Studio registry's web search tools (Parallel SDK when set; keyless MCP fallback).
`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.