Phidata

Agno's pitch is unusually concrete for the category: the SDK and runtime are open source and free at any scale, and the vendor explicitly does not meter events, tokens, compute, storage, retention or egress. That structure only works because the platform runs in your cloud, on-premise or locally — the trade is that you carry the infrastructure bill and the ops work yourself. What you buy is the control plane: $150/mo billed monthly for Pro, which points the console at one live AgentOS connection with 3 seats, basic role-based access control and email support; Enterprise adds end-user management, end-user audit logs, custom RBAC with per-resource scoping, SAML SSO and Slack Connect support. Strengths are in the build-and-run surface. Three primitives — agents, teams, workflows — cover most orchestration shapes without forcing a graph DSL, and the runtime is framework-agnostic, so LangGraph, DSPy, Claude SDK or hand-written code can run side-by-side with Agno agents. Serving is genuinely multi-surface: REST for your product, MCP for Claude and ChatGPT, interfaces for Slack, Discord and WhatsApp. AgentOS ships durable execution, distributed state, request isolation, resumable streams, websockets, background execution and A2A alongside JWT/service-account RBAC, audit logs and built-in tracing. The v3.0.x changelog shows engineering attention where it counts: incremental session-history loading keeps long conversations from slowing down, tool JSON schemas are derived once and cached, RecursiveChunking no longer duplicates trailing chunks, embedding failures during knowledge ingestion now surface as partial files with retries instead of silent empty vectors, and MCPConfig(stateless=True) lets any replica answer any request. Claude extended-thinking blocks now replay verbatim, and Gemini receives each image's real MIME type instead of a hard-coded JPEG. Weaknesses are mostly about scope and orientation. The Free plan is real and unlimited, but its control plane is local-only — live operating sits behind Pro. Complex hand-built graph workflows and state machines are the domain of LangGraph-style tools; Agno offers agents, teams and workflows rather than an arbitrary graph editor. The framework is Python-first and assumes a container-capable environment: if you cannot run Docker, a cloud account or a Kubernetes cluster, the ownership model becomes a burden rather than a benefit. Budget-sensitive small teams should price self-hosting honestly before assuming $0. Where it fits: internal agent platforms, customer-facing copilots and support agents you need to keep inside your own perimeter, Slack-native agents with approval gates, and document or research workflows where your data must never leave your database. Where it doesn't: teams that want a fully managed hosted service and no infra responsibility, and teams already deeply committed to LangGraph's ecosystem who would gain little from swapping runtimes.