A problem solver, task automator, and question answerer. Ask a question; it breaks it down, matches each piece to real endpoints, and tells you honestly what can run today. Every verified run strengthens the route, so the next person's answer is faster, cheaper, and proven.
first 50 asks free on any key
One question in → its definitive sub-questions out, with what-to-know ranked and the blanks only you can fill named explicitly. Recursive: zoom into any branch.
Every leaf is matched to real endpoints and labeled runs-now, plannable, or blocked. Nothing is claimed executed unless it actually fired and passed its gates.
Verified routes are stored, weighted by real outcomes, and reused, including by paraphrase and by class ("speakers" learns from "headphones"). Usage makes it smarter.
# REST, ask a question (X-API-Key from your welcome email)
curl -X POST https://liminality.physea.ai/reduce \
-H "Content-Type: application/json" -H "X-API-Key: $KEY" \
-d '{"question": "how do I send a weekly digest email from git activity?"}'
# every response is structured JSON with a legend explaining itself;
# blanks name the slots only you can fill before executing.# MCP, wire any model straight into the engine (remote, no install) claude mcp add liminality --transport http https://liminality.physea.ai/mcp \ --header "X-API-Key: $KEY"
Using Claude Code or another MCP client? Add the line above and the full toolset (lookup_library, reduce_question, ground_question, decide_question, …) is available remotely over this same API, your key meters tool calls just like REST (the handshake is free). A local stdio install also ships with the engine.
Self-serve: pay at checkout and your API key is emailed automatically to the address on your receipt, usually within a minute, no waiting on a human. Packs are 20, 100, or 1,000 asks; the first 50 asks on any new key are free. Questions:
Popularity never overrides verification. A well-worn answer still has to pass its checks before it is reused.