Introducing AG-UI; the Agent-User Interaction Protocol.
AG-UI is an open, lightweight, event-based protocol that standardizes how AI agents connect to fronted-end applications.
After a year of one-off, CopilotKit <> Agent frameworks collaborations, we decided to ship a standard, and open way for agents to connect to apps.
We are launching with day 0 integrations with LangChain, Mastra, CrewAI and AG2 with more partnerships on the way.
How AG-UI works:
- Creates a unified, two-way bridge between backend agents and frontend apps
- Works with any event transport (SSE, WebSockets, webhooks)
- Supports 16 standardized event types for common agent-user interactions
Where AG-UI fits in the agent protocol landscape:
- MCP: for agent←→tools communication
- A2A/ACP: for agent←→agent communication
- AG-UI: for agent←→human interaction in applications
We launched the protocol yesterday, and have been blown away by the reception. We have the first working group this Friday, to help expand and steer the direction of the protocol.
People use the same model / server for all queries not because it's sensible, but because it's simple. This brings the same simplicity to the far more optimal solution.
And great startup play too, by definition no incumbent can fill this role.
We are building CopilotKit = a framework + platform for building context-aware AI assistants into any application (not necessarily coding related applications).
Part of CopilotKit is about giving the Copilot / AI agents access to the application through a typed "inline realtime API". And to make ergonomics great for _our_ users, CopilotKit ships with hardcore type programming.
Cursor's Copilot (unrelated to CopilotKit) helped us write this type code (that's what went viral). And yes, GitHub Copilot is a mis-attribution.
By the way - we practice what we preach-
Copilots raise the bar (or the ceiling...) on human productivity - in every domain. Which is why we're building infrastructure to make building copilots easier...
What ideas do you think are still missing from today's Copilots?
I.e. suppose we were looking back at today's Copilots 5 years from now- besides better models, what else has changed?
I wrote the article (not the code - though I was part of it).
The code is is 30 lines of conditional, recursive, partially-type-inferred Typescript type expression. It's actually not too bad when you break it down.
What it enables: Zod-like safety + convenience, over plain JSONs.
Three experiments looking into how GPT4, Gemini & LLaVA perform when getting conflicting visual and textual information. Some pretty interesting results.
Shameless plug: I work on CopilotKit - open-source copilot building blocks for react apps.
Designed to alleviate exactly the pain points in the article.
Devs define simple Copilot entrypoints--
state (frontend + backend + 3rd party),
action,
purpose-specific LLM chains, etc. And the CopilotKit engine takes care of the rest.
There's an interesting opportunity to _meaningfully_ "open-source the algorithm" by just letting each person define their own curation, at least on top of the provider's initial curation.
The internet of that world looks different from our world, probably in better ways for consumers.
AG-UI is a launch partner of A2UI, but it is a separate project by CopilotKit, not google.
We have a day-0 handshake between AG-UI & A2UI