Agents Frame
OpenRouter routes models. Agents Frame routes minds.
Agents Frame is the thinking-framework router for AI agents. POST an intent, get the right cognitive framework — first principles, inversion, second-order, 26 more — picked by semantic recall + ELO reranking. REST + MCP + CLI.
one intent · three frames · three answers
The answer your agent gives depends on the frame it picks. Agents Frame returns three so you can see the question from angles a single model would have collapsed into one.
Strip the question to its primitives.
What does this system need to do this quarter? Forget the codebase that exists. Sketch the minimum that satisfies that — then compare what's there to what's needed.
→ Refactor only the pieces the new shape still needs.
Past investment is not an argument.
The 80,000 lines from last year are spent. Whether to keep them is a forward-looking question about future maintenance cost — not a question about respecting prior work.
→ Discount the existing code to zero. Decide from there.
What follows the obvious answer?
Refactor saves three weeks now. But what state will the team be in six months later? Will anyone want to touch this module again? Pick the option whose downstream effects you want to live with.
→ Optimize for the team you'll have, not the team you have.
pipeline
Four deterministic steps. No LLM judge in the hot path. Latency stays predictable so you can put it inside agent loops without budgeting an extra 5-second tail.
Auto-route en / zh / de / es / fr / pt — the embedding step needs the right tokenizer or recall collapses.
~8ms
1536-dimension vector through gemini-embedding-001. Pinned model, deterministic dim, cached for repeated intents.
~210ms
Top-K candidates pulled from a curated library of 15+ thinking frameworks indexed in Postgres + pgvector.
~40ms
Each framework carries an ELO score updated by user feedback. Rerank, return the top three, ship a ready-to-paste prompt.
~170ms
three entry points · one router
Same routing engine, three ways to call it. Pick the one your stack already speaks — switch later without rewriting prompts.
POST any intent, get three frameworks back.
curl -X POST https://api.agentsframe.com/v1/think \
-H "Authorization: Bearer $AGENTS_FRAME_KEY" \
-H "Content-Type: application/json" \
-d '{
"intent": "Should I rewrite or refactor this legacy code?",
"topK": 3
}'pricing
Five free calls a day, no card. Two paid tiers when you need more — metered by routing calls, not user seats. Cancel anytime from the billing portal.
Five framework lookups a day. Forever, no card.
For builders shipping multi-step agents weekly.
For teams running agents in production.
ready when you are
One framework lookup before each PR review, one before each design spike, one before each tough call. That's a week.
Should I rewrite or refactor this legacy codebase?
Ask what failure looks like first.
What state will the team be in 6 months later?
Past LOC is not an argument for keeping them.