Feature Request: Add AG-UI Protocol Support

[anirudha]

What would you like to be added?

HolmesGPT currently supports CLI and Robusta UI. I’d like to propose adding support for the AG-UI protocol — an open, event-driven standard for agent–user interaction.

AG-UI streams structured JSON events (e.g. TEXT_MESSAGE_CONTENT, TOOL_CALL_START, STATE_DELTA) over SSE or WebSockets, allowing agents to plug into modern UIs and frameworks like CopilotKit or LangGraph with minimal glue code.

Proposed Implementation

• Add a new output mode that streams HolmesGPT responses as AG-UI events
• Map:
  • Agent output → TEXT_MESSAGE_CONTENT
  • Tool calls → TOOL_CALL_START / TOOL_CALL_END
  • State updates → STATE_DELTA
• Default transport: SSE (with optional WebSockets)
• Config flag: --ag-ui (off by default for backward compatibility) or a new http server ?

---

Open Questions

• Should HolmesGPT start with output-only events, or support bi-directional AG-UI immediately?
• Should we design tool-specific schemas (e.g., Prometheus, K8s), or stick to generic event wrappers?
• Should this be part of core HolmesGPT or an external adapter?

---

Next Steps

• Prototype SSE streaming of TEXT_MESSAGE_CONTENT
• Gather community feedback on event mappings

@ashwin-pc @goyamegh

---

References

• [AG-UI Protocol Spec](https://github.com/ag-ui-protocol/ag-ui)
• [Introducing AG-UI Blog](https://webflow.copilotkit.ai/blog/introducing-ag-ui-the-protocol-where-agents-meet-users)

### Why is this needed?

- Improves interoperability with AG-UI compatible front-ends   
- Makes HolmesGPT’s reasoning and tool usage visible in real-time  
- Increases adoption by developers already using AG-UI in their ecosystems  

[aantn]

Hi @anirudha,
Sorry on the delay in replying!
I love this and it makes a lot of sense.

Regarding your questions:

1. I don't have a strong preference on output-only vs bi-directional. Although probably its simpler to start with output-only.
2. Can you explain this more? (I didn't see anything in https://docs.ag-ui.com/ about tool-specific schemas, so curious if there is a precedent for this)
3. I would do it in core HolmesGPT. I think it'll be clunky to do it outside of HolmesGPT

Regarding other implementation details, I know @moshemorad wants to refactor some of our internals (specifically ToolCallingLLM so there is a better split between the core agentic functionality and the client/view layer). Once that work is done, I suspect it would be easy to add support for AG-UI in the view layer. (I'm using 'view' in a broader sense here, referring to ui/cli/http-api. It's basically the internal users of the core agent class.)

[anirudha]

ack, a side question /
is there ongoing effort or does homesgpt run as a server ?

[aantn]

It does! See https://holmesgpt.dev/reference/http-api/

[aantn]

Note that I think there is a bug in the docs on how to specify the model when using the HTTP server. @pavangudiwada is working on a fix!

[pavangudiwada]

@anirudha The HTTP server docs are now fixed https://holmesgpt.dev/reference/http-api/