[Feature]: Implement AG-UI to ACP Jaeger AI (PoC)

[Nabil-Salah]

Requirement

Requirement

Implement a Proof of Concept (PoC) for the Jaeger backend to act as a stateless protocol translator bridging the frontend UI and an external AI sidecar. The backend must accept connections using the AG-UI protocol (from CopilotKit) and proxy those requests to an external "Bring Your Own Agent" (BYOA) sidecar using the Agent Client Protocol (ACP) over WebSockets/application (stdin/stdout).

Context & References

• Epic/Tracking: AI-Powered Trace Analysis with Local LLM Support: - AI-Powered Trace Analysis with Local LLM Support #7832
• AG-UI Protocol: Docs
• Agent Client Protocol (ACP): Docs
• Backend PoC PR: feat: Add agent-client-protocol PoC: -feat: Add agent-client-protocol PoC #8196
• Frontend Poc PR: -[feat]: AI chat (CopilotKit) + AG-UI integration jaeger-ui#3664

Architecture & Request Flow

• https://docs.google.com/document/d/1-HZejtvQy2Edv7sQNp8mSzx2B2tgpOOSOKiRs4KN0b0/edit?usp=sharing

Implementation Tasks for PoC

1. Gateway Routing & Auth

• Create the jaegerai package and register the /api/ai/chat HTTP handler.
• Ensure the handler utilizes existing Jaeger tenancy/bearer token middleware.

2. Protocol Translation (Go Backend)

• Define the Go structs required to parse incoming AG-UI requests.
• Implement the ACP WebSockets adapter to dial the external sidecar (e.g., ws://localhost:9000).
• Build the streaming loop: Read incoming ACP notifications from the sidecar and format them as AG-UI Server-Sent Events (SSE) for the HTTP response writer.

3. Test Environment (The Sidecar)

• Provide a lightweight Python sidecar script using the official agentclientprotocol SDK and Gemini API to act as the test target for the Go gateway.

Problem

• AI-Powered Trace Analysis with Local LLM Support #7832

Proposal

No response

Open questions

No response

[Nabil-Salah]

while working to make this translation layer, i went into a problem
the ag-ui prompt -> include content text, ToolCalls (tools are sent with every single request, the frontend only sends the tools that are relevant to the exact components currently rendered on the user's screen) -> so it's related to every prompt not every session
for acp prompt -> payload is primarily just text
and tools can be defined per session
so for this how i can map?
What I was thinking of is if an ACP client (jaeger) wants to give an ACP agent (sidecar) some tools, it spins up an MCP server, and tells the agent to connect to it.
This MCP will update its list of tools with every prompt using notifications/tools/list_changed  here
@yurishkuro So what do you think? can this acceptable?

[yurishkuro]

why would MCP tools change from conversation to conversation? Tools should be stable. If you're thinking of different tools depending on different context (problem being solved) then this could ideally be handled via skills hierarchy, not the changing of tools configuration.

[Nabil-Salah]

This MCP is different from the backend MCP,
the extra one is the dynamic ones are strictly for driving the user interface tools available from the frontend and it changes with every prompt request.
Those tools usually include things like describing what is the user view, toolboxes, navigateTo

In the CopilotKit/AG-UI paradigm, tools aren't just backend functions; they are also React UI components. Because the frontend is stateful and components mount/unmount as the user navigates, the tools available to the agent must reflect exactly what is currently rendered on the user's screen.

[yurishkuro]

PTAL https://docs.google.com/document/d/13BZDc3gFDP8zp89nfixssDZLPlvnV0v-bl6sb_j03ts/edit?tab=t.0

[Nabil-Salah]

yes that's a good approach

• it will lower token consumption
• lower hallucination from big contexts

[abhay1999]

Went through the design doc — using skills hierarchy instead of changing tools per prompt makes sense. Fixed tools, but the agent pulls in different knowledge depending on what it's debugging.

One thing I'm not sure about: if the first message is vague like "checkout is slow", how does the agent pick where to start in the skill tree? Does the root skill file handle that routing, or is it just left to the agent to figure out?