Agno

What it is

Agno is an open-source platform for building agents and running them as services. You write agents with the Agno SDK, or with the Claude Agent or Codex SDKs. AgentOS then runs them behind an API and an MCP server, with login, scheduling and a record of every run. This listing covers the part that touches your coding agent. A template repository, a setup skill and an MCP connector for the docs let Claude Code, Codex or Cursor stand up the platform and build agents on it.

1. You paste one prompt into your coding agent
2. It clones a template and starts AgentOS with PostgreSQL
3. You build a first agent together
4. You connect the AgentOS UI, then the API or MCP
5. You deploy it and schedule or improve its agents

Who it suits

Developers and small teams who want agents running as a real service, with a login, history and a schedule, instead of a script on a laptop. It also suits anyone who would rather have a coding agent set the platform up than wire it by hand. The SDK is Python.

If you only want one script, install the SDK alone and skip the template. The docs call a script fine as a one-off.

Where to use it

Agno’s docs give full steps for three coding agents, and a general pattern for the rest.

• Claude Code: the setup prompt, a connector that lets it read the docs, and agno connect for your own platform
• Codex: the same three, with codex mcp add for the docs connector
• Cursor: the docs connector in ~/.cursor/mcp.json, and agno connect
• Claude Desktop: agno connect writes a local config entry that starts the mcp-remote bridge through npx, so Node.js must be installed
• Claude and ChatGPT: a custom connector to your platform on a public HTTPS URL. A secured endpoint must use OAuth, and availability varies by provider and plan
• Any other agent that adds MCP servers: the docs give a general pattern (name agno-docs, URL https://mcp.agno.com, Streamable HTTP, no authentication)
• A coding agent without MCP: the https://docs.agno.com/llms.txt and https://docs.agno.com/llms-full.txt indexes

The platform runs in Docker on your machine, or on a cloud. The README lists templates for Docker, Railway, AWS, GCP, Azure, Fly.io, Render, Modal and Helm. Agents can also answer in Slack.

How to set it up

Get it saves a zip of Mindket’s setup notes in your downloads folder, with each agent’s steps from Agno’s docs. The main route is one prompt, pasted into your coding agent.

1. Paste the prompt below into your coding agent.
2. Let it run the setup skill. It starts AgentOS and PostgreSQL and checks the API and MCP connection. Add your model credentials to the local environment file when it asks.
3. Open os.agno.com, sign in and choose Connect OS. The overview page uses Local and http://localhost:8000.
4. Build your first agent with the prompts below.

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

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

This is the Build your agent platform page’s prompt. The README gives much the same prompt with agentos-railway, a template that deploys to Railway.

To let your coding agent read the docs while it builds, add the docs as a connector. For Claude Code, the Add Agno page gives this line. For Codex it gives codex mcp add agno-docs --url https://mcp.agno.com.

claude mcp add --transport http --scope user agno-docs https://mcp.agno.com

How to use it

The template ships skills that your coding agent runs. Name one in your prompt so the agent knows to use it.

Then use what you built in the AgentOS UI. You chat with an agent, open its sessions and traces, and see the searches and page reads behind an answer. To call an agent from your own app, the Run your agents page has you create a service account in the UI with the agents:read and agents:run grants. Set its token as AGENTOS_API_TOKEN and your platform’s address as AGENTOS_URL. Then send a form request to the agent’s runs endpoint, as the page does for Daily Brief.

curl --fail-with-body "$AGENTOS_URL/agents/daily-brief/runs" \
  -H "Authorization: Bearer $AGENTOS_API_TOKEN" \
  --form-string "message=Generate today's edition." \
  --form-string "stream=false"

With stream=false the answer comes back as JSON with its status, content, run ID and session ID. To continue the conversation, send the same session ID with the same token.

Example

You have the platform running and want a daily briefing on your field. You send the docs’ prompt, changing the interests to your own:

Build an agent called Daily Brief.

I work on AI developer tools. Follow agent frameworks and platforms,
model and coding-agent releases, and practical examples of agents
running in production. Prioritize developments that could affect
what we build, how we explain it, or who might use it.
Skip generic AI commentary.

Use the Parallel MCP to research the web.
Produce a three-piece edition with one lead story and up to two
other developments. For each, explain what changed, why it matters
to me, and link to dated original sources. Prefer developments since
the previous edition; use the last 24 hours for the first edition.

Save previous editions in the shared notes.
Use per-user learning to remember each user's preferences.

Your coding agent creates the agent in agents/, registers it in app/main.py and adds a description and sample prompts to app/config.yaml. You refresh the Agents list in AgentOS, open Daily Brief and ask it to generate today’s edition. The prompt asks for a lead story and up to two other developments, each with what changed, why it matters to you and dated source links. The session and trace show the searches behind it.

For newcomers and experts

If you are new. An agent is a language model with tools and instructions. A script that runs one agent is fine for a first try. AgentOS keeps an agent running as a service instead. It answers over HTTP, stores sessions and memory, and traces each run, so you can see what happened. The docs suggest starting with a Daily Brief, an agent that researches what you follow. Let your coding agent do the setup. If something is wrong, give it the example and ask it to fix the behavior.

For experts. AgentOS can serve Claude and Codex agents next to Agno’s own. The README says you build with the Agno, Claude Agent or Codex SDKs, and AgentOS adds authentication, role-based access, resume and replay, human approvals and durable execution. When the AgentOS has authorization on, agno connect mints one service account per coding agent, with a 90-day expiry by default. You can revoke one with agno tokens revoke claude-code without touching the rest. An unsecured AgentOS connects without credentials, and an OAuth-protected MCP endpoint gets a one-time sign-in instead.

The failure cases are in the docs. With the default max_attempts of 1, a crashed run can fail visibly. More attempts restart from the original input and can repeat external actions, so the guide to running Claude and Codex as services says to make state-changing tools idempotent. For several instances, the scaling page adds Redis for live run events and cancellation, and says to deploy that with one instance before raising the count.

Good to know

Most of Agno runs where you put it, and a few paths reach out.

• Where it runs: the README says to run the platform in your own cloud and keep your data in your control. The AgentOS UI at os.agno.com connects from your browser to your running platform
• Usage data: Agno sends a telemetry event per agent so it knows which model providers to prioritize. The README says prompts, messages and outputs are never sent. Turn it off with AGNO_TELEMETRY=false
• Model calls: agents call the provider you configure with your key. The Daily Brief example also searches the web through the Parallel MCP
• No security by default on Render: RUNTIME_ENV=dev leaves the deployed platform open to anyone with the URL. Set it to prd and turn on JWT authorization before you add private data
• Coding agents: the Claude and Codex engineers in the template are off by default. Once on, they run on the API host with its environment and credentials, until you set up sandboxes
• Data loss: agno down --volumes also removes the named volumes, which destroys data
• Removal: the pages Mindket read give no uninstall steps for the SDK or the template. agno disconnect removes the AgentOS entries from your coding agents’ configs. Add --revoke to also revoke the service accounts agno connect minted. That needs the running AgentOS and an admin credential
• License: Apache-2.0