Files read properly.
Prompts spoken aloud.
Scanned pages are found and read on your yes, right-to-left text keeps its order, and you can speak a prompt instead of typing it.
Turning a file into text is half the job. The other half is knowing what the text can be trusted for. AigoPilot never passes off a page it could not read, labels anything a model wrote, and keeps the original next to the text. The token side of this, files and web pages turned into clean text, has its own page.
Every file you attach keeps three things side by side in the project's folder: the original, the text taken from it, and your note about what it is. Write the note in the box under the file, or let AigoPilot take it from the prompt you send the file with. The agent is told all three, and is told exactly what is missing from the text, such as pictures of pages or hidden sheets.
PDFium, the PDF reader inside Chrome, returns right-to-left lines in the wrong order: a number in the middle of a Persian sentence comes out with its parts reversed. AigoPilot rebuilds every line from the position of each letter instead, puts brackets and half-spaces back, and reads two-column pages one column at a time. A Persian sample PDF came out word for word correct.
A page that is only a picture of text, or whose letters cannot be decoded, is found by itself and never passed off as read. One click and the model reads it with your own Claude subscription; you can say yes for later files too. The text it writes is labelled as written by the model, with a reminder to check every number against the original, and the tokens it cost are shown.
The page is first turned the right way up. In a real test, two scanned pages, one of them upside down, were both read without a single wrong character. You can also install Tesseract and read scans on your computer instead.
Press the microphone beside the paper clip and speak. By default the voice is written down on your own computer by whisper.cpp, in the language you set (Persian by default), and only put into the box where your cursor is. Nothing is sent until you send it. The microphone is opened only while you record, and silence or noise never turns into made-up text.
In the settings you can use OpenAI's speech engine with your own key instead. The price per minute is shown, and a budget stops it before anything is sent. English technical words inside a Persian sentence are still the weak spot: on Persian programming sentences about a third of the words came out wrong, so read what it wrote. The same voice typing is in the iPhone app, which is still in testing. We have not yet tried it with a real microphone.
AigoPilot gathers the names a project uses from its own files (its folder, package.json, Cargo.toml, requirements.txt and its git branches) and gives them to the speech engine as a hint. A project word that comes out in Persian letters is written back in Latin. The list is a plain text file you can edit, and a word you remove is never gathered back.
A microphone beside the search box takes short spoken requests: open a project by its name, find a session by what was said in it, list the sessions waiting for you, or read the last answer aloud. It only moves around the page. It never writes to a session or answers one. Built, but we have not yet tried it with a real microphone.
Say an instruction, confirm it by voice, hear the summary of the answer, and approve a permission only with a spoken yes that was read back to you first.
Skills are checked for security before they are installed and after every update, tested on their own first, and installed for Claude Code and Codex once you approve.
When a session does something well, it can write down how it did it. After a security check and a test, that becomes a skill every project can use.
Keep your ideas in one place, and turn one into a project with its first session when you are ready.
Planned for a later version.
Yes. It finds pages that are pictures of text, and on your yes the model reads them with your own Claude subscription, or Tesseract reads them on your computer if you install it. Text written by the model is labelled so you check the numbers against the original.
PDFium, the PDF reader inside Chrome, returns right-to-left lines in the wrong order, so a number inside a Persian sentence comes out reversed. AigoPilot rebuilds each line from the position of every letter instead.
Not by default. It is written down on your own computer by whisper.cpp, and the text is only put into the box. You send it yourself. If you choose OpenAI's speech engine in the settings, the recording goes to OpenAI with your own key.
