Introducing Replicate's new search API to find the best models (replicate.com)

Replicate launched a new search API (available in the latest TypeScript alpha and Python pre-release) that returns ranked results for models, collections, and documentation pages for a given query. Model results include the full model object (url, description, run_count) plus a new metadata object with tags, a longer generated_description, and relevance scores. Example responses show rich metadata — e.g., google/nano-banana with run_count 5,426,257 and a generated_description describing multi-image fusion, targeted edits, and SynthID watermarking. The endpoint is reachable via HTTP or the SDKs as replicate.search(), and docs are in the HTTP reference and OpenAPI schema. This is significant because it makes model discovery and selection far more developer-friendly and LLM-integrable: MCP servers (remote and local) already support the API and perform dynamic jq-style response filtering to keep LLM context windows small by returning only the most relevant fields. The search API integrates with tools like Claude Desktop, VS Code, Cursor, OpenAI Codex CLI, and Gemini CLI, and the TypeScript SDK includes type hints for easier development. The old search endpoint is disabled in MCP in favor of this beta API; Replicate asks for feedback as the feature evolves.