Agno

• Building a customer support agent that answers from product docs and order data, then serving it inside your product over the AgentOS REST API
• Running a team of research agents that fan a question out to specialists and merge the answers into one brief
• Publishing an internal agent as an MCP server that employees add to Claude or ChatGPT as a custom connector
• Scheduling a weekly workflow that summarizes product usage and posts the report to Slack
• Adding approval gates to an invoice-processing workflow so a person signs off before anything is posted
• Standardizing on one runtime, database and control plane for agents written in Agno, LangGraph and DSPy

Agno's distinguishing choice is that the production runtime is part of the open-source package rather than a paid hosting tier. AgentOS gives you the API server, database-backed sessions and memory, JWT auth, approvals, tracing and scheduling, and Agno charges only for the hosted UI that watches live deployments. That makes it attractive for Python teams that would otherwise build that platform layer themselves and want data to stay in their own cloud. The costs are operational: you own the deployment, database and network security, releases move fast and sometimes need migrations, and the most useful control-plane features for a team start at $150 a month. Pin versions, read the release notes before upgrading, and turn off telemetry if your policies require it.

About Agno

Agno is the renamed Phidata project: the old Phidata docs now say "Phidata is now Agno!", and the phidatahq/phidata GitHub repository redirects to agno-agi/agno. It has three layers. The Agno SDK (`pip install agno`, Python 3.9+) gives you three primitives, agents, teams and workflows, plus attachable capabilities such as tools (100+ prebuilt toolkits and MCP), knowledge for agentic RAG, user memory, session state, skills, reasoning, structured input and output, human approval steps, evals and guardrails. Models are set with a `provider:model_id` string, and the docs cover native adapters for OpenAI, Anthropic, Google Gemini, Mistral, Cohere, DeepSeek, xAI and others, cloud platforms such as AWS Bedrock, Azure and Vertex AI, gateways such as OpenRouter and LiteLLM, and local runtimes such as Ollama, vLLM and LM Studio. AgentOS, the runtime, is part of the same open-source package. It turns your agents into a production API with SSE and WebSocket streaming, stores sessions, memory, knowledge and traces in your own database, adds JWT-based RBAC with optional per-user data isolation, OpenTelemetry tracing, cron scheduling and background runs, and exposes agents through an MCP server, A2A, AG-UI, Slack, Telegram, WhatsApp and Discord. It can also serve agents built with LangGraph, DSPy or the Claude Agent SDK through adapters. Starter templates deploy it as a container to Docker, Railway, AWS, GCP, Azure, Fly, Render, Modal or Kubernetes, and Agno's quick start has you paste a setup prompt into Claude Code, Cursor or Codex. The paid part is the hosted AgentOS Control Plane at os.agno.com, a web UI for chat, sessions, trace trees, a no-code Studio for versioning components, memory, knowledge, evaluations, approvals and schedules. The browser talks to your AgentOS directly, and Agno says it never holds a copy of your data. The Free plan connects the control plane only to AgentOS instances on your own machine; Pro ($150 per month) adds one live deployment connection, three team seats and roles; Enterprise adds end-user identity, audit logs, per-resource RBAC and an optional self-hosted control plane. The tradeoffs: you operate the runtime, database and network controls yourself, the SDK is Python-only, the project ships breaking changes quickly (v3.0 in August 2026 required a database migration), and per-run telemetry is on until you disable it.