Wiring AGNT memory into Claude Code through MCP

Claude Code's MCP client is the cleanest example we've seen of how tool-calling should work in an IDE-native agent. Declare the server in your config, and every tool it exposes becomes callable from the agent without writing any wrapper code.

AGNT has exposed its memory + wiki + graph layer as an MCP server since v0.4. The endpoint lives at `/mcp/sse` and currently surfaces `search_venues`, `list_venues`, and `confirm_booking`. The next surface drop adds `read_user_memory`, `write_user_memory`, and `query_knowledge_graph`.

Why this matters for Claude Code builders: if you're scripting workflows that touch user context ("what's this user's dietary restriction?", "what venues did they book last month?", "what's in the AGNT knowledge graph for Canggu?"), you no longer need to shell out to curl or write a custom adapter. Claude Code calls the MCP tool, AGNT returns structured JSON, and the reasoning step happens inline.

Configuration is three lines in your Claude Code MCP config. We're shipping a full walkthrough in /guides/ecosystem/wire-agnt-memory-into-claude-code with the exact JSON, the API key setup, and the scope guardrails you should set on the token.

What not to do: don't treat the MCP endpoint as a dumb REST wrapper. The whole point of MCP is that Claude gets structured tool schemas — use the tool-calling abstraction, not raw HTTP, so the model can plan multi-step queries without you hardcoding the sequence.