May was a milestone month. We crossed 40,000 GitHub stars, which would be worth writing about all on its own, but it landed alongside some of the most consequential releases we've shipped. The v2.6 line reshaped how agents reach external data, brought other frameworks under the AgentOS runtime, gave agents safe access to the local filesystem, and made Claude prompt caching far more granular. A big month, and a foundational one.

Here's what stood out.

🎉 Milestone: 40,000 GitHub stars

We hit 40,000 stars this month. It's easy to wave off stars as a vanity metric, but behind this one are thousands of people filing bugs, opening PRs, and putting Agno into real projects. That's the part we care about. Thank you.

New releases & features

The v2.6 series turned Agno's external-data story and its multi-framework ambitions into first-class APIs. Here are the highlights:

Context providers (`agno.context`)

The headline of v2.6.0: a first-party API for plugging an external source (filesystem, web, SQL database, Slack, Google Drive, or an MCP server) into an agent as a natural-language tool. Instead of hand-rolling a retrieval pipeline for every source, you attach a context provider and the agent reasons over it directly. This is the foundation the rest of the v2.6 providers build on.

See the Context Providers docs.

Multi-framework support in AgentOS (beta)

AgentOS now runs more than just Agno agents. A single AgentProtocol backbone gives basic support for ClaudeAgentSDK, Langgraph, and DSPy on the AgentOS runtime, so teams can bring existing agents under one operational surface instead of standing up a second stack.

Learn more about Multi-Framework Support in our docs.

Team HITL, approvals, and resume runs

Human-in-the-loop graduated to teams in v2.6.0: a full API layer for Team HITL plus Approvals, both wired into the AgentOS chat page. Workflow executor-level HITL also landed for cases where a pause-tool flow is set up on an agent or team inside a specific workflow step. And background runs over SSE can now be reconnected and resumed in AgentOS, so a refresh mid-run picks up where you left off instead of starting over.

Dive into our HITL docs.

Factories for runtime creation

AgentFactory, TeamFactory, and WorkflowFactory, added in v2.6.0, let you build agents, teams, and workflows dynamically at run-time, unlocking cleaner multi-tenant patterns where each request spins up its own configured instance.

View the Dynamic Agents docs for more.

Claude prompt caching (multi-block)

v2.6.1 brought three Anthropic prompt-caching enhancements scoped to the Claude model: system_prompt_blocks with per-block cache and optional ttl ("5m" or "1h"), a cache_tools flag that caches the tool prefix (Anthropic, Bedrock, VertexAI), and deterministic tool ordering in Model._format_tools so request prefixes stay stable across runs and caches actually hit.

View the Prompt Caching docs for more.

Parallel MCP backend

v2.6.1 added ParallelMCPBackend, a web backend for WebContextProvider that talks to Parallel's public MCP server and exposes web_search + web_fetch with compressed markdown output. Keyless by default, with Bearer auth via PARALLEL_API_KEY for higher rate limits and an optional OAuth path.

View the MCP Parallel Agent docs for more.

Workspace tools

v2.6.2 added a polished local-machine toolkit giving agents read / list / search / write / edit / move / delete / shell access to a root directory tree, with destructive operations gated behind Agno's built-in human-in-the-loop confirmation by default. The safe way to let an agent actually work on a project on disk.

View the cookbook.

WorkspaceContextProvider

v2.6.3 added WorkspaceContextProvider, a project-aware context provider for repository roots. It's backed by the read-only Workspace toolkit rather than generic FileTools, so agents get a view of a project that understands its structure.

Learn more in the Workspace docs.

Gemini multimodal file search

v2.6.5 added multimodal support to the Gemini File Search API, available with google-genai≥1.75.0. See the cookbook example for an image-upload walkthrough.

Gmail and Calendar context providers

Also in v2.6.5: GmailContextProvider and CalendarContextProvider, both aligned with the existing provider patterns. GDriveContextProvider also gained OAuth support alongside its existing service-account auth, so an agent can connect to a user's own Drive without a service account.

Learn more about the Gmail Provider and Calendar Provider in the docs.

Slack HITL: multi-row approvals

v2.6.6 brought multi-row approvals to the Slack interface, with all pause types supported, so a reviewer can resolve several pending approvals in one place without leaving Slack.

Learn more about Slack HITL in the docs.

Gemini interactions

v2.6.7 added a GeminiInteractions model class built on Google's new stateful interactions API, so agents can use the interactions endpoint directly.

View the Gemini Interactions docs for more.

Per-user data isolation in AgentOS

Also in v2.6.7: an opt-in per-user data isolation layer for AgentOS authenticated endpoints, so a single deployment can keep each user's data partitioned.

Learn more about per-user data isolation in the docs.

Google Antigravity support

v2.6.8 brought first-party support for Google's Antigravity API in two shapes. AntigravityAgent is a BaseExternalAgent served through AgentOS with native sessions, streaming, and UI. AntigravityTools is a toolkit that lets any Agno agent delegate a sub-task to a managed Antigravity sandbox.

Learn more in the Antigravity docs.

Gemini managed agents

Also in v2.6.8: GeminiInteractions support for Google's managed agents. That covers Deep Research, which does autonomous research with citations and background streaming with reconnect and resume, and Antigravity, a general-purpose agent running in a managed Linux sandbox.

See the Gemini Interactions docs for more.

Community projects & showcases

Here's some of what people built with Agno in May:

MindRoom: federated, decentralized agents by basnijholt

An Agno-powered workspace where AI agents live in Matrix rooms, remember context, use tools, and work across bridged chats like Discord, Slack, and WhatsApp. Multi-agent, multi-user, and federated by design.

Check it out on GitHub.

TaskBounty: agents that earn USDC by eliottre

A marketplace where autonomous agents compete on real tasks and get paid in USDC/ETH/SOL on approval. Escrow-based, with a REST API and OpenAPI 3.1 spec, plus a new agent-native referral mechanic, Bounty Scout, where agents earn credits by referring clients.

Check it out at task-bounty.com.

Secure Playground: prompt-injection attack & defense simulator by Harish Kotra

An interactive prompt-injection simulator built on a FastAPI + Agno multi-agent pipeline, with Agno evaluator agents scoring attacks and defenses in real time. Stack: Next.js, TypeScript, Tailwind, Framer Motion, Zustand, FastAPI, SQLAlchemy, and WebSockets.

See the writeup on LinkedIn.

AgnoLab: a visual canvas for Agno by devalexandre

A visual canvas interface for building and orchestrating Agno agents. Think of it as a playground where you can watch agent workflows run.