July reshaped how you run Agno. v2.7.0 brought AgentOS to the command line with service accounts, a CLI, a focused MCP surface, and eval suites that keep regressions out of production. Then v2.8 added rollouts that turn successful runs into training data, a durable filesystem, and an ops agent that lets you ask questions about your own platform in plain English.

Let's take a closer look.

New releases & features

Connect any AgentOS to your coding agent

The headline of v2.7.0. A new CLI ships as agnoctl on PyPI and runs as agno. Run uvx agno connect and it discovers a running AgentOS, mints a per-client access token, writes the MCP config for Claude Code, Claude Desktop, Cursor, Codex, and ChatGPT, and verifies the handshake, so you skip hand-editing JSON. Those tokens are service-account personal access tokens (agno_pat_...) built for real security: stored as SHA-256 hashes, scoped per user, expiring after 90 days by default, revocable on demand, and enforced across every deployment mode including /mcp and WebSockets. From there, agno tokens create/list/revoke manages tokens, agno create scaffolds a project from the agentos-<provider> templates for Docker, AWS, Fly, GCP, and Railway, and agno up/down/restart/status handles the local lifecycle.

See the Connect Your Clients docs for more.

MCP Interface v2

v2.7.0 gave the /mcp endpoint a clean eight-tool operator surface: get_agentos_config, run_agent, run_team, run_workflow, continue_run, cancel_run, get_sessions, and get_session_runs. It trims run results by default so responses stay lean, sends progress notifications on long-running tools, and carries the full human-in-the-loop continue and cancel lifecycle, so a coding agent gets a predictable, purpose-built way to run and manage AgentOS instead of a sprawling tool list.

See the AgentOS as MCP Server docs for more.

Eval suites and CI gating

Also in v2.7.0: a new agno.eval layer with a proper suite runner built from Case and run_cases/arun_cases, plus an argparse CLI that supports team subjects and numeric judge scoring. SuiteResult.to_dict() is a stable JSON contract you can wire straight into CI, so you score agents and teams consistently and fail a build on regressions instead of eyeballing results.

See the Evals docs for more.

Unified auth across every transport

v2.7.0 put a single AuthMiddleware in front of REST, /mcp, and WebSocket transports, so authentication no longer drifts from one to the next (JWTMiddleware stays as a backward-compatible alias). Scope enforcement now runs identically across JWT, service-account, and MCP paths, A2A and AGUI routes are authorized rather than just authenticated, and concurrent MCP requests are isolated so one can't pick up another's state. In v2.7.2 you can also put standards-based OAuth in front of /mcp with AgentOS(mcp_auth=...), so clients authenticate the way the rest of your stack expects.

See the AuthMiddleware docs to learn more.

Turn your agents' best runs into training data

v2.8.0 added a straight path from evaluation to fine-tuning data. agno.scorer grades attempts, agno.environments runs an agent over a set of tasks in full isolation, and run_rollouts runs each task K times so you measure a real pass rate (pass@k) instead of whether it worked once. to_sft_jsonl then exports the attempts that passed as conversational SFT data, with a provenance sidecar tracing every example back to the run it came from. Each attempt runs on a fresh db, session, and user, so no state bleeds between runs to contaminate your pass rate or your training set. If all you want is scoring inside an existing suite, Case.scorer drops any scorer into a Case with no extra model call.

See the cookbook for more.

A durable filesystem your agents can rely on

The filesystem in most setups is a scratch directory that vanishes when the run ends, which is fine for a one-shot task and useless for anything an agent is supposed to remember. v2.8.2 added FileSystem, a durable text store an agent writes to and reads back across runs, so a decision it recorded or a checkpoint it saved is still there next time it starts up. You choose where files live, whether SQLite for dev, Postgres for a multi-worker app, or local disk to open them yourself, without changing agent code. Templated namespaces like assistant/{user_id} scope each user's files to their own space, resolved from the run context at call time, so the model can't reach into another namespace by passing a different path.

See the FileSystem docs to learn more.

Ask your AgentOS how it's doing, in plain English

v2.8.5 added AgentOSTools, a read-only ops view of the AgentOS an agent runs on. Point it at your database and it reports on usage, latency, failures, schedules, evals, components, and pending approvals, then answers questions like which tool was slowest today or how many runs failed this week, grounded in real traces instead of a dashboard you assemble by clicking around. It reads from the database rather than a live AgentOS handle, so an ops agent can inspect the whole platform with no path to change anything: it watches, it doesn't touch.

See the cookbook for a working example.

Run agents and knowledge on Valkey

v2.7.3 brought Valkey to Agno as an in-memory backend. ValkeyDb holds session and state for agents, teams, and workflows in memory, so reads and writes stay quick under load when latency matters more than durable long-term storage. Valkey also lands as a vector store that runs both vector and keyword search from the same backend, giving you hybrid retrieval over one in-memory store instead of standing up a separate keyword index alongside your vector database to get it.

See the Valkey database docs and Valkey vector store docs for setup.

Persistent sandboxes with Superserve

v2.7.4 added SuperserveTools, which lets an agent write and run its own code inside a Superserve sandbox. The sandbox is a Firecracker microVM, and the part that matters is that it persists: files the agent writes and packages it installs are still there on the next tool call and the next run in the same session, which is the difference between running one-off snippets and building something up over many steps. Secrets are handled through a proxy token, so agent-written code can call real APIs without your credentials ever landing somewhere the model can read them.

See the Superserve toolkit docs to learn more.

Production observability with The Context Company

Also in v2.7.4: an observability integration with The Context Company. Call instrument_agno() before you import Agno and your runs stream over OpenTelemetry, capturing model calls, tool arguments and results, prompts, token usage, and latency. Where plain tracing shows one run at a time, this analyzes across all of them, surfacing recurring patterns and account-level insight and linking back to the specific runs behind each one, so instead of scrolling traces you find where agents consistently go wrong and go straight to the evidence.

View Agno’s The Context Company docs to learn more.

Run frontend tools from AG-UI

v2.7.2 added client tool execution to AG-UI, so a tool can run in the frontend rather than only on the server. Your UI can handle actions that belong on the client, like reading local state or driving the interface, while the agent orchestrates the flow.

See the AG-UI docs for more.

New tools and integrations

July added a stack of new toolkits and integrations:

• TwelveLabsTools (v2.6.22): analyze videos and generate multimodal text embeddings, so video becomes something an agent can search, summarize, and reason over. See the TwelveLabs docs.
• SofyaTools (v2.6.22): search, extraction, and research through a single integration. See the Sofya docs.
• SearchApiTools (v2.6.22): Google, News, Images, and YouTube search from one toolkit. See the SearchAPI docs.
• RedmineTools (v2.7.3): find, read, create, and update Redmine issues, leave comments, and log time from an agent. See the cookbook for a working example.
• TokenLab (v2.7.3): a new OpenAI-compatible model provider you can point agents at with the interface you already know. See the cookbook for a working example.
• PlivoTools (v2.7.4): send SMS, place voice calls, and look up a number's carrier and line type before sending. Check out Agno’s PlivoTools docs to learn more.
• SmallestTools (v2.8.6): text-to-speech with Smallest AI's Lightning models, output as WAV, MP3, or telephony formats. Check out Agno’s Smallest AI toolkit docs for more.
• OpenSearch (v2.8.6): a vector database with vector, keyword, and hybrid search in both sync and async variants. See the cookbook for a working example.‍
• LocalFileSystemTools read (v2.6.21): read files, not just write them, scoped to your target_directory by default. See the Local File System docs.

Community projects & showcases

Our community built incredible things in July. Here are two standout projects:

whatbroke: a CLI diff tool for agent behavior by Arthi Arumugam

Agents drift. A tool call drops, an argument changes, latency spikes, and you don't notice until it's too late. whatbroke diffs agent behavior between two runs and surfaces exactly what changed: tool calls, arguments, outputs, cost, and latency. If you trace your Agno agents with Langfuse, you can export from there and run the diff.

View the project on GitHub.

Social Media Terminal: multi-agent content automation by Muhammad Ikhwananda Rizaldi

An advanced CLI workflow powered by Agno that automates the whole content lifecycle, from campaign brief to deep research to QA to auto-publishing. A full content department in a terminal.

See the post on LinkedIn.

Contributor shoutouts

Community MVPs

• @basnijholt: Cached the Pydantic version lookup during tool wrapping, which was reading package metadata on every single wrap and stalling the event loop for multiple seconds on network-backed installs. The fix dropped benchmark times from 65.94ms to 11.01ms for 100 wraps (#9210).
• @harshitajain165: Shipped the Smallest AI text-to-speech toolkit, synthesizing speech and returning it as an audio artifact in a ToolResult, with optional disk saving (#9015).
• @jperla: Added TrustedRouter as a first-class model provider, a thin OpenAILike subclass pointing at the TrustedRouter API, with cookbook and tests (#9100).
• Rushikesh Gandhmal: Fixed nested-team memory continuity, where delegated teams could lose conversation context mid-turn even with history enabled. Delegated teams now remember what came before (#8968).
• Anshul Jain: Integrated You.com's Research API into Agno, giving agents deep-research capability aimed at real compliance and logistics problems where a missed deadline costs far more than a fine.

More great contributions

• @khrnchn: Preserved explicitly configured Agent and Team member IDs exactly as provided, fixing a mismatch where an ID like billing_agent was exposed one way but resolved another during delegation, breaking handoffs (#9011).
• @RayST3: Updated the Moonshot integration for kimi k3, adding use_thinking handling, round-tripping assistant reasoning_content, and new cookbook examples (#9057).
• @alec-drw: Fixed skill access tools so a model calling get_skill_script or get_skill_reference without the path no longer trips a ValidationError before the graceful handling can run (#9096).
• @pratikm778: Serialized nested executor requirements so StepRequirement.to_dict() safely converts nested RunRequirement objects for standard JSON persistence (#9162).
• @devdattatalele: Fixed six broken links across the cookbook READMEs, repointing the docs links to the reorganized /knowledge/concepts/ paths (#9195).

Every one of you made Agno better. Thanks for building with us.

Get involved

There's never been a better time to contribute.

• ⭐ Star Agno on GitHub
• 💬 Join our Discord
• 📖 Read the docs
• Check out open issues
• 🚀 Share what you're building. We want to feature YOUR project next month.

— Team Agno