Add agentic chat system with MCP tooling

[MagellaX]

Summary

• wire the new MCP-backed agent runner into chat routing with feature-flag control
• expose the dedicated MCP tool server for conversations, memories, and action items
• include regression tests and e2e harness for agent streaming

covers #2935

[MagellaX]

hey @Arshroop-Saini wdyt? any takes?

[Arshroop-Saini]

@MagellaX I already built this, check PR #3048

[MagellaX]

@MagellaX I already built this, check PR #3048

Nice work on #3048 direction’s solid!! That said, a few bits feel verbose/brittle: a global app_state with cross‑module import from main, lots of print/emoji logs, graph compiled in two places, a string-based tool loop guard, and a metadata-only vector “full scan” that’ll be pricey. My PR delivers the same agentic behavior but cleaner and lower‑risk: MCP tools + agent flow behind a feature flag with automatic fallback, tools/graph stored on FastAPI app.state (no globals), slimmer prompts/logging, DI instead of importing main, and full unit/E2E coverage with streaming + agent action logging. I mean, it’s a smaller, surgical diff that stays backward compatible and is easier to operate/monitor. I’m happy to fold in anything missing, but merging this version should cut noise, reduce regressions, and keep iteration fast.

[beastoin]

@MagellaX interesting.

btw, why do we need the mcp when we could build function_tool(s) to achieve better performance?

and, do we have any test results, sir?

[MagellaX]

@MagellaX interesting.

btw, why do we need the mcp when we could build function_tool(s) to achieve better performance?

and, do we have any test results, sir?

hey, thanks for digging in. I kept the MCP layer because we already expose the same 14 tools (memories, action items, Pinecone search, etc.) to the wearable, the CLI, and now chat.
By going through MCP ,we only maintain the tool logic once, we can gate/rotate credentials server-side, and we get consistent usage/telemetry across surfaces. If we were to re-implement them as function tools inside LangGraph, it would duplicate the adapters, drift from the other clients, and require touching multiple stacks for future tool additions(which can be done, but would take up a bit of time). The streaming hop adds only a few ms on my profiling, so the practical performance cost is tiny compared with the reuse we get.

[beastoin]

#3092

[beastoin]

@MagellaX @Romamo would be great if you guys could collaborate and help omi chat great // also feel free to ping me @thinh via discord http://discord.omi.me