Ask anything. Each answer lights the path it took through the diagram above. When Oracle7 cannot verify an answer it says so and says why, and the Gaps tab can turn those declines into the specialists your model is missing.
Point Oracle7 at something of yours and it becomes a reader you can ask about it. Each answer names where it read the value. Everything stays on this computer. Type /help in the chat for every feature and command.
1Scan this computer
System, CPU, memory, disks, uptime and what is using the most memory. Figures and process names only, never your files.
2Connect a server
Read a server over SSH with your own key: hostname, uptime, load, memory, disks, top processes, failed services. Read-only; no password is ever asked for.
3Add files
Drop tables (.csv, .tsv) or text (.txt, .md) here, or . Ask "what is the price of item 4?" or "what do my notes say about the boiler?".
4Connect an MCP server
Give the command that starts it or its URL. Pick its tools and the Workbench builds a specialist that routes each question to the right tool, shows the exact call it made, and learns from your right / wrong marks. Questions routed to a server are sent to it.
5Connect a sensor
A device or hub that serves JSON or metrics over HTTP (ESP32, Shelly, Tasmota, Home Assistant, a weather station), an MQTT topic, a USB/serial device, or a log file. No hardware yet? Try the simulated garden node. Readings are kept for 24 hours on this computer; answers give the value, its age and its source.
Or write a specialist by hand on the Build tab.
No sensors yet.
+Connect a sensor
A specialist is a small program that answers one kind of question and declines the rest. It joins your model only by passing its held-out exam; it can only add answers where the core declined.
With the journal on, each question you ask and whether it was answered is kept in workbench/local/journal.jsonl on this machine only. Nothing is sent or packaged. Declines are grouped into the specialists your model is missing.
Ask for a function. The code organ hands over only programs it has run, so it writes what it can prove and declines the rest. It can write and prove: a small working website, palindrome check, factorial, vowel count, word count and word frequencies, HTTP routes, regular-expression extraction, memoised recursion, a bank-account class, matrix multiply, fizzbuzz, fibonacci, reverse a string, JSON round-trip, concurrent work, binary search, tower of hanoi, quicksort, sum a list, prime test, a number-guessing game, tic tac toe. The bench runs any code against tests of your own in a separate Python process with a time limit.
An essay on a topic, composed from this model's own records. Every sentence is a sentence from its sources, kept word for word with its number, unless you turn on rewording: then a sentence may be reworded, and only when reading it back gives exactly its source's meaning (each one is listed with its source). ProseFold chooses which sentences to use and how to join them, to fit the length, reading level, depth and style you pick. Nothing is made up, and nothing leaves this computer. How well the essays read has not yet been measured by people.
Record answers as a film: the diagram tracing each path, the question, the answer and its receipt. Made in your browser as a .webm file; nothing is uploaded.
Ask something first; each answer can be filmed.
+Narration, sound and images optional · ElevenLabs
Connect your own ElevenLabs account to narrate each answer, lay a sound you describe underneath and open each answer with an illustration (images need their Pro plan). This is the only feature that sends anything off this computer: the narration text, sound description and image prompts go to ElevenLabs with your key. Generated images are labelled as generated in the film.