Get live chat progress

[jm8985]

I hope this isn't a silly question, but I'm building a UI that interacts with LLMs / mcp servers and I would like the ability to chat with agents programmatically. I wasn't able to test the FastAPI example, on startup it throws: AttributeError: 'Namespace' object has no attribute 'message'
How can I grab what's displayed in the console and pass that to a custom function?

[jm8985]

I've tried running fastagent as an mcp server, and referencing it in my mcp client app. The agent has a rather long task, and I don't see any status updates or messages for a few minutes, until the agent completes the task. Is there some other means to get progress messages back during execution?

[jm8985]

This project is ground breaking, and honestly super exciting, but it's missing a few key ingredients. I want to write an MCP client in the same vein as cursor, cline, Claude desktop, etc. But it's not possible, as once an agent receives a message, it goes off on its own and does not respond programatically until after its task has been completed. Granted, I can see what's happening in the console, but I have no way of exposing that to the user. I could grab the logs, but there is no session ID. It would be a huge benefit to everybody if there were a FastAPI integration that supports streaming of all events (reasoning, tool calls, tool results)
Can I get a hear hear, or some acknowledgement or criticism?

[jlyst]

I also would like access to this data for the same reasons. Similarly I want access to the completion data coming from the provider. I am using openrouter and want to track token usage ($).

[evalstate]

As a stop-gap until it's built in to the API, the OpenTelemetry reporting is very rich.

[yalattas]

I spent few hours trying to figure this out
I tried to integrate fastapi with FastAgent but the main problem was getting FastAgent responses programmatically
I was able to initialize the agents and configured a couple of MCP servers
Yet, when I push a prompt to FastAgent, logs shows MCP in running state but never responded for more than 15 minutes

This project is extremely good and resolve many issues
Yet, without programmatic support for another applications to use it. FastAgent will be good as decorations more than practical usage

If I understood something wrongly, Appreciate correcting me

[evalstate]

Is there any code you can share? Quite happy e.g. to look at PRs that "fail" and figure out what needs to be done (or whether this would be supported).

It should be possible (I might add an example later) to programmatically create an Agent using an AgentConfig (NB plan to rename that to AgentCard in line with the A2A protocol). I also plan to bring the Tool Stop/Call loop closer to the Agent rather than within the LLM classes.

[jm8985]

Sorry to keep pushing, but the elephant in the room is that without basic streaming access to the agent, this project is handicapped.

[evalstate]

What would be the ideal API for this? Would this be yielding out intermediate request/response/tool call pairs or the underlying token stream?

[jm8985]

Yielding out request/response/tool call pairs would be nice. However, one of my agents generates html reports (llama_index ReActAgent), and I've written a wrapper around the token stream to yield custom events to cleanly pull out content in a UI. This way, my system prompt can instruct the LLM to wrap content to be extracted with xml tags. That would be a pretty neat trick if fast-agent could support that.

[evalstate]

I wasn't able to test the FastAPI example, on startup it throws: AttributeError: 'Namespace' object has no attribute 'message'

Whicḣ example was that?