Five source formats ingest into one library.
Resources takes the material your team did not write and makes it machine-readable: paste a URL and the server strips the nav and the scripts, or upload a PDF up to 25 MB and pdf-parse pulls the text out. Search then runs over titles, summaries, tags, and the full extracted body. No embeddings yet.
Resources attaches a source to Gmail by slug.
One ingested URL, PDF, or EPUB that every granted AI Employee can search, export, and attach to a Gmail draft.
One paste turns a web page into text
The server fetches the page and extracts readable text, stripping scripts, nav, and footers, with no browser or scraping stack required. A failed fetch keeps the row and its error so a human can fix it.
PDFs, EPUBs, and Markdown up to 25 MB
PDFs extract via pdf-parse, EPUBs unzip chapter by chapter, and TXT, Markdown, and HTML upload directly, up to 25 MB per file and 1 MiB of extracted text each.
One query searches titles, tags, and full bodies
Search runs over titles, summaries, tags, and extracted bodies. Humans type and watch the results narrow. AI Employees query that same surface through a tool.
EPUBs open in a reader that remembers your progress
Detail pages know the type they hold. Text is editable markdown, PDFs use the native viewer, EPUBs get an in-app reader with table of contents and saved progress, and a URL gets an open-original card.
Chromium renders the PDF, so tables survive
Export any Resource as PDF, HTML, Markdown, or plain text, with PDFs rendered through Chromium so headings, tables, and code blocks come out styled for a chat reply or a Base record.
Gmail attaches a Resource by its slug
Gmail send and draft tools accept attachments by Resource slug. The server checks the Grant and resolves the bytes, so no base64 ever crosses the model's context window.
Mira cited the billing guide in her Run.
AI Employees reach the library through built-in tools gated by three Grant levels — read, edit, delete. The tool descriptions coach them to check whether the team already ingested a primer before improvising.
They curate it too
An employee can file a URL, a pasted transcript, or a file it already holds with create_resource — the PDF a customer emailed, or a Word contract it converted first. It gets full control of rows it authored, while teammates start at read.
Levels, not switches
read covers list, search, and get; edit adds re-titling, tagging, and body updates; delete allows permanent removal. Humans promote employees between levels from the share modal.
New material is instantly usable
Every new Resource is granted read to all AI Employees, and every employee you hire is granted the library that already exists — so the primer you drop in at 9:00 informs the Routine that runs at 9:05, and the employee you hire next month reads it too.
Readers ask 5 questions about Resources.
How is a Resource different from a Note or a Memory?
A Resource is content the team did not write — an article, ebook, or transcript ingested once and queried on demand. A Note is a page the team authors together, and a Memory is a durable fact auto-injected into an AI Employee's prompt.
What formats can I ingest?
Web pages by URL (fetched and extracted to plain text), PDF, EPUB, TXT, Markdown, and HTML uploads up to 25 MB per file, and pasted raw text. Video files are accepted but transcripts aren't extracted yet — upload the transcript as text in the meantime.
Can AI Employees add their own Resources?
Yes. The create_resource tool lets an employee index a URL, file a pasted transcript or research summary, or file an actual file it already holds — a PDF a customer emailed over, a Word contract it converted to PDF, a form it downloaded. The authoring employee automatically gets full control of its own row; teammates start at read-only. Video files still need a human, because transcripts aren't extracted yet.
Can an AI Employee email a Resource to someone?
Yes. The Gmail send and draft tools accept attachments by Resource slug and format — the server checks the employee's Grant, resolves the bytes, and attaches the original file or the text rendered as PDF, HTML, Markdown, or plain text.
Does it use embeddings or RAG?
Not yet, and retrieval no longer depends on it. Search matches every word of a query independently and in any order across titles, summaries, tags and the full extracted text, then returns the matching passage and its position — so an employee reads the paragraph that answers the question rather than the first chapter of the document that contains it. Embeddings and vector search are still planned once real query patterns are known.
4 more products share one database.
Bases
Multi-table workspaces with 11 field types, saved views, comments, and attachments, plus 21 built-in tools for a granted AI Employee.
Notes
Markdown pages in nested notebooks, with read or write Grants that cascade from a notebook down to every sub-page.
AI Employees
Teammates that carry a markdown Soul and work Routines on a schedule you pick, with every Run transcribed and costed.
Workspace
Channels, DMs, and 25 MB file uploads where AI Employees are members you can @mention.
Tomorrow one thing is finished before 09:30.
Install Genosyn, register an AI Model, write one role, and put it on a schedule. The company grows from there.