Say it once. Land it right.
Deterministic fork detection runs on every keystroke, in your browser — zero API calls. If your text supports more than one reading, each fork surfaces with its exact span, the readings it supports, and a one-line clarifier that locks the one you meant. Clean input renders nothing.
You're reading this page in the Technical register — the “Reading as” switch in the header changes it. Same facts, read your way.
Free. No account. Your words never leave the page until you send them.
Type a prompt like you'd say it out loud. If it can be misread, the readings appear here before you send it — then one tap opens your own AI app with the clear version filled in.
Or watch it catch something — try one:
Nothing is sent anywhere until you pick an app and hit send there yourself.
Bring the reply backpaste any AI answer — read it the way you talk
1
Type like you talk
No format constraints — raw, fragmented input is the expected case, not an error.
2
See what you said
Detection is structural and deterministic: identical input, identical output. Only real forks surface; clean prompts stay silent by contract.
3
Send the one you meant
Pick a reading and its clarifier is appended; one tap opens your own model app with the clarified prompt filled in. The call is never proxied.
Both directions
Misreads travel both ways. So does this.
For agents
The same engine, pointed at your AI.
Agents misread workloads the way people misread messages. The engine under the box also derives runtime settings — and names the failure mode it's preventing.
The research behind it — including the misses.
Every recommendation traces to a law that was checked against criteria fixed before the data was generated. Three registered checks failed during development — all three are reported in full, because a scorecard you can trust has to include the misses.
The research & the full scorecard →Your notes
Your notes, in every AI you use.
Your Obsidian vault — or any folder of notes — becomes memory your AI can actually use, without leaving your machine. Nothing is readable until you choose folders; every answer shows exactly which notes it used; and anything the AI saves lands back in your vault as an ordinary note, linked to its sources, visible in your graph. Works in Claude Desktop, Claude Code, Cursor, and anything else that speaks the open connector standard (MCP).