Quick reference

One endpoint: POST /v1/messages. Required: model, max_tokens, messages.

Always check stop_reason before using content.

Retry 429, 500, 529 with backoff; fix 400, 401, 403, 404, 413.

Batch: 50% off, up to 100,000 requests or 256 MB, 24-hour expiry, match by custom_id. Cache: prefix match in order tools, system, messages; 5-minute default TTL; writes 1.25x, reads about 0.1x; max 4 breakpoints.

Tools: great descriptions; all parallel results in one message; errors as is_error: true; strict: true for schema-valid inputs.

MCP: tools (model), resources (app), prompts (user); stdio local, Streamable HTTP remote. Security: prompts are not boundaries; least privilege, isolation, approvals, hooks. Choose: workflow before agent; Skill teaches how, MCP gives access; pin model versions.