Status: idle
This session history is persisted with a Durable Object per session ID.
1) Tools / Platform / Stack
- Platform: Cloudflare Workers (Python runtime)
- Model: Cloudflare Workers AI (
@cf/meta/llama-3.1-8b-instruct)
- Tooling: MCP via
MultiServerMCPClient
- MCP Servers: Cloudflare Docs + AWS Knowledge
- Frameworks/Libraries: LangChain + LangGraph + langchain-mcp-adapters
2) Chat Design (Architecture)
- Worker route
POST /api/chat receives user prompt.
AgentRuntime runs a LangGraph flow: planner -> tools -> finalize.- Planner decides model-only or MCP tool path.
- Tools node runs one or more MCP tools and collects links.
- Finalize node returns structured response + provenance metadata.
- Session context is keyed by client cookie (
sid) and persisted in a Durable Object for cross-request continuity.
3) User Flow
- User enters prompt and clicks Send.
- Status bar updates: request -> planning -> tool/generation.
- Agent executes graph with session history context.
- Assistant returns
User Intent + detailed Answer (+ Resources when available).
- UI shows provenance path/tool usage for transparency.
4) Future Expansion
- LangSmith tracing: add node-level trace and latency visibility.
- Security hardening: prompt-injection guardrails, output filtering, and tool allowlists.
- Auth & policy: role-based access and per-tool permission controls.
- Observability: structured logs, error taxonomy, and token/cost dashboards.