AgentOS is a durable agent runtime that serves agents over API, MCP, and chat interfaces like Slack. Build agents for your product, your team, or yourself. Run them in your own cloud, keep their data in your own database, and manage them from AgentOS.
Built with Agno, this template gives you a working agent platform that you can run locally or deploy to Modal (+ Neon Postgres).
Three ways to build agents.
- Use a coding agent. Open the repo in Claude Code, Codex, or Cursor. The included skills help it create, test, and improve your agents.
- Describe what you need. Chat with OS Manager, the built-in agent that helps you manage the platform and build new agents.
- Use Studio. Build agents visually in AgentOS Studio.
Three ways to use your agents.
- In your product. Call your agents through the REST API.
- In AI apps. Connect apps such as Claude and ChatGPT through MCP.
- In Slack. Connect a Slack app to talk to OS Manager from your workspace.
The AgentOS UI lets you chat with your agents, review past conversations, and see the steps an agent took. Your PostgreSQL database stores conversations, memories, and Company work.
Copy this prompt into your coding agent:
Help me set up my agent platform and build my first agent.
Clone https://github.com/agno-agi/agentos-modal into a folder called agent-platform,
cd into it, and follow .agents/skills/setup-platform/SKILL.md.
It checks Docker, helps you add your API key, starts the platform, and walks you through building your first agent. Already cloned the repo? Ask your coding agent to follow setup-platform in that folder.
Prefer to set it up yourself? Follow the steps below.
The platform includes Claude Engineer and Codex Engineer. They can read your code, make changes, and run commands.
Chat with either engineer in AgentOS, or ask OS Manager to use one:
Ask Codex Engineer to review this project and suggest improvements. Don't change any files yet.
Codex uses the OPENAI_API_KEY you added during setup. To use Claude, add
ANTHROPIC_API_KEY to .env. After changing .env, run docker compose up -d again.
By default, engineers work in this project's folder. With Docker Compose, they can change files in your local checkout, including the code serving the platform. Review those changes before committing them.
For engineers to work in a separate environment, configure a sandbox with your own repository URL and branch:
- Local Docker sandbox: run the API on your computer and engineering work in a separate Docker container.
- AWS AgentCore sandbox: run engineering work remotely when the platform is deployed. The API host needs AWS credentials and an S3 bucket for checkpoints. No disk is attached to the API service.
Both engineers stay available with either setup. Sandboxes are optional and are not enabled by the default Compose configuration.
You'll need Git, Docker with Compose, an OpenAI API key and an Anthropic API key for Claude Engineer. Make sure Docker is running. Run the commands below in a terminal.
git clone https://github.com/agno-agi/agentos-modal agent-platform
cd agent-platform
cp example.env .envOpen .env and set OPENAI_API_KEY and ANTHROPIC_API_KEY so both engineers can
authenticate. For Claude Code OAuth instead, remove ANTHROPIC_API_KEY and set
CLAUDE_CODE_OAUTH_TOKEN. Then start the platform:
docker compose up -d --buildThe first build can take a few minutes. When it's ready, open http://localhost:8000/docs to see the API.
Check that AI apps can connect through MCP:
./scripts/mcp_check.shThe check should report 9 tools; required tools available.
- Open os.agno.com and sign in.
- Choose Connect OS, select a local connection, and enter
http://localhost:8000. - Name it Local AgentOS, then open the chat for OS Manager.
Tell OS Manager what you want to build. For example:
Help me build an agent that answers questions about my product. Here's the documentation URL: …
Review what it proposes, then ask it to build and publish the agent. Refresh the Agents list and try it out. You can chat with your new agent directly or ask OS Manager to run it for you.
You can keep working through OS Manager, use Studio, or edit Python files in this repo. The repo includes instructions, called skills, for common coding tasks. Ask your coding agent to follow the relevant skill:
| What you want to do | Skill |
|---|---|
| Build an agent | /create-agent |
| Add a tool, change behavior, or fix a bug | /extend-agent |
| Write tests for your agent's answers | /create-evals |
| Harden an agent with simulations and targeted fixes | /improve |
| Check the repo before sharing or releasing it | /review-and-improve |
If you write an agent in Python, put it in agents/ and add it to
app/main.py so the platform can run it. Set its description and
suggested prompts in app/config.yaml. To make a tool or model
available in Studio, add it to app/registry.py.
You can give agents documents to search, a notebook to keep notes, and memory to remember each user. Keep a product's documentation in its own knowledge base, separate from the platform's documents.
Company organizes ongoing work. A Product is what you're working on, a Project groups related work, and a Task is one piece of work to complete. A Routine is a set of instructions that can run on a schedule.
This template includes an AgentOS Product for improving the platform itself. To get started, ask OS Manager:
Review this codebase and recent Agno releases. Suggest the most useful improvements.
It also includes a daily review at 09:00 UTC. Company Manager reads the code and Agno release notes, then proposes up to three Tasks for you to review. The review is instructed to suggest work and leave approval and execution to you. Automatic acceptance is off.
Scheduled reviews use your model API keys and incur usage costs. Check the saved schedule and run history in Company to see when a review will run and what it did.
app/company.py sets the initial review instructions and schedule. Once they're saved in the database, update the saved Routine to change them; editing the Python file alone won't change an existing Routine.
To save your work in your own GitHub repository, create an empty private repo.
From a fresh clone of this template, run the commands below with your repo URL.
This makes origin your repo and keeps the original template as upstream:
git remote rename origin upstream
git remote add origin https://github.com/YOUR-ORG/YOUR-REPO.git
git push -u origin HEADFuture pushes go to your repository. You can still fetch updates from upstream
and review them before merging.
If you use sandboxes, also set the repository URL and branch in the sandbox configuration.
Ask your coding agent to follow deploy-platform, or use the steps below.
You'll need Modal CLI with modal token new completed, Neon CLI with neonctl auth completed, Python 3, and OpenSSL.
[ -f .env.production ] || cp example.env .env.productionAdd OPENAI_API_KEY and any other model keys you use, such as ANTHROPIC_API_KEY
or CLAUDE_CODE_OAUTH_TOKEN for Claude Engineer. Keep RUNTIME_ENV=prd or leave
it unset. The scripts read .env.production before .env and upload every key
in the file to the agentos-secrets Modal secret.
Set NEON_ORG_ID in .env.production (find it with neonctl orgs list). To use an existing database, fill in the DB_* settings instead. An existing database is not deleted by the teardown script unless its NEON_PROJECT_ID is recorded.
Creates a Neon PostgreSQL database and a Modal secret, then deploys modal_app.py. The image is built from this checkout and runs in one always-warm container.
./scripts/modal/up.shThis creates or uses billed resources. Start with one API instance; see the scaling guide for the optional multi-replica architecture. When the script shows the public URL and asks for a verification key:
- Open os.agno.com, choose Connect OS, and select Live.
- Enter the public URL and enable token-based authorization.
- Copy the full public verification key into the script's prompt.
The script saves the key and connection settings in .env.production. If you
skipped the prompt, save the quoted JWT_VERIFICATION_KEY there and run
./scripts/modal/env-sync.sh to finish. Production requires a valid token
to access agents and data.
Open /health and /docs on your public URL, then connect through the AgentOS UI.
Check that /agents rejects requests without a token and works with a valid token.
Use the provider dashboard to review startup logs.
Use your public URL followed by /mcp as the connector URL in an AI app.
When prompted, enter the MCP_CONNECT_SECRET saved in your production env file.
For a supported coding client, run:
uvx agno connect --url https://YOUR-DOMAINTry asking OS Manager about the platform to check that the connection works.
| What you want to do | Command |
|---|---|
| Deploy code changes | ./scripts/modal/redeploy.sh |
| Upload production settings | ./scripts/modal/env-sync.sh |
| Remove the deployment | ./scripts/modal/down.sh |
Removes the Modal app and managed secret, and deletes the Neon project when NEON_PROJECT_ID is recorded. Review the resources listed before confirming deletion.
The API service needs no attached disk. Conversations, memories and Company work are stored in PostgreSQL. Without a sandbox, engineers run inside the API container; their edited files and native sessions can be lost on redeploy or replacement. Database conversation history does not reconstruct that working directory.
For durable, isolated engineering, configure AWS AgentCore and S3. S3 stores the last saved workspace and native-session checkpoint. Recovery does not preserve unfinished edits at the instant of a crash or replay external actions safely. AWS is optional for getting started. Local Docker isolation is also available in the sandbox guide.
Before running multiple API instances, follow Scaling.
To check your Python code and test agent responses, install uv
and run these commands from the project folder. Keep the local platform running.
If another Python environment is active, run deactivate first.
./scripts/venv_setup.sh
source .venv/bin/activate
./scripts/validate.sh
python -m evals --tag smokevalidate.sh checks Python formatting, lint rules, and types. The smoke tests ask
OS Manager to explain setup, check platform health, and explain agent memory.
They use your model API keys and save results in the database for review in the UI.
PostgreSQL must be running, even when you only want to list the tests.
Add your own cases in evals/cases.py, or use
/create-evals. To save a report, add
--json-output tmp/evals.json. To run one case, use --name CASE_NAME.
Evals use your application's database and workspace; agents can call tools during a test.
Change dependencies in pyproject.toml, then run ./scripts/generate_requirements.sh.
Commit the updated requirements.txt and rebuild the API and any sandbox image you use.
See AGENTS.md for the code layout and development conventions.
example.env shows the main settings. Start with OPENAI_API_KEY;
add the others when you need them. Use .env locally and .env.production for deployment.