// the take
A thin wrapper layer around @anthropic-ai/claude-agent-sdk that handles session lifecycle, message normalization, and WebSocket transport (Node and Bun) so you don't rewrite that plumbing for every Claude-powered chat UI. It's aimed at people building their own Claude Code-style web or desktop clients, not end users.
Clean interface boundary (IClaudeAgentSDKClient) between the session layer and the actual SDK call, so you can swap in retries, logging, or a different transport without touching Session/SessionManager. It solves a genuinely annoying problem: reading and resuming from the .claude/projects JSONL transcript files on disk, which is undocumented SDK-adjacent behavior most people would have to reverse-engineer themselves. Ships two complete, runnable reference implementations (Bun WebSocket server, Express+Vite+ws server) rather than just snippets, and includes AbortController-based interrupt support for in-flight streams, which is easy to get wrong.
Pinned to @anthropic-ai/claude-agent-sdk@^0.1.0 with a README note telling you to manually bump the version yourself — that's a sign the SDK is still moving fast underneath this and breakage on upgrade is likely. Session persistence is just local JSONL file reads from .claude/projects, so it doesn't work across multiple server instances or survive a container restart; the 'swap it for a database' extensibility point is aspirational, not implemented. There's no authentication or authorization anywhere in the WebSocket handler — any client that can open a socket connection can drive an agent session billed to your ANTHROPIC_API_KEY, which is a real problem the moment this leaves localhost. The project itself calls some pieces 'upcoming UI primitives,' meaning parts of the advertised surface aren't built yet.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →