​
What is Agents Frame?

Agents Frame is a thinking-framework router for AI agents. Give it an intent (“debug a flaky CI test”, “decide whether to refactor”, “ship a launch plan”) and it returns the best-fit mental model with a ready-to-use prompt — already filled with the framework’s reasoning scaffolding. Where OpenRouter routes you to the right LLM, Agents Frame routes you to the right way of thinking. The router is language-aware (en/zh), runs Gemini embeddings against a curated framework library in pgvector, and re-ranks candidates with an ELO-weighted scorer.

Quickstart

Get an API key and make your first /v1/think call in under 60 seconds.

Architecture

How the routing pipeline works: detect → embed → recall → rerank.

​
Three entry points, one routing engine

The same router is exposed through three surfaces. Pick the one that fits how your agent (or you) likes to talk.

REST API

POST https://api.agentsframe.com/v1/think — Bearer auth, JSON in/out. The lowest-friction path for any backend.

MCP Server

Streamable HTTP at https://api.agentsframe.com/. One tool: route. Drop-in for Claude Desktop, Cursor, and any MCP host.

CLI

npm i -g agents-frame then agents-frame route "<intent>". The fastest way to feel the router from your shell.

​
At a glance

  • Library: 15 frameworks live across thinking / business / psychology categories; growing toward 30. Each is bilingual (en/zh).
  • Pricing: Free 5 calls/day · Starter 19/mofor1,500calls⋅Pro19/mo for 1,500 calls · Pro 29/mo for 3,500 calls. Yearly saves 17%.
  • Latency: ~300–600 ms end-to-end for a typical /v1/think call. Detect → embed → cosine recall → ELO rerank, all in one request.
  • Stateless: No conversation state on the server; you keep your agent’s memory, we just answer “which framework?”.

​
When to reach for the router

Pre-flight a tough decision

Before letting an agent commit to a destructive action (ship, delete, refactor, fire a customer), POST the intent and let three frameworks argue first.

Diversify a single-model answer

Running one LLM tends to collapse the answer to one frame. Get back three to sanity-check the obvious one.

Standardize team prompts

Stop hand-rolling prompts per problem class. Each framework ships with a tuned scaffold — paste and go.

Audit-trail an agent

Every call is logged with intent, candidates, ELO scores, and feedback. Replay why an agent picked the framework it did.

​
When not to use it

  • Pure retrieval / RAG — wrong tool. Use a vector DB directly.
  • Zero-shot question answering — call the LLM, save the round trip.
  • Sub-100ms hot loops — the embedding step adds ~210ms. Cache framework picks per intent class instead.

​
SDKs and surfaces

  • REST — works in any language with an HTTP client.
  • MCP — drops into Claude Desktop, Cursor, Windsurf, and any host that speaks the Model Context Protocol.
  • CLI — npm i -g agents-frame. Pipes well into git hooks and PR review scripts.
  • Official Node SDK is on the V1 roadmap. For now, every language has REST.