How Globus works

Step one · connect

Each person connects their own account

There is no company-wide crawl and no administrator who hands over a shared drive. Access is granted one person at a time, by that person, and everything built from it stays keyed to their address.

Signing in

A 6-digit code, no password

An employee enters their work email and receives a code. A code is only sent if the domain is registered to this organisation, and the match is exact equality — not a suffix test, so a lookalike domain does not resolve.

There is no SSO, no SAML and no directory sync. Email codes are the whole of it today.

What they link

Google, and only Google

On the corporate portal an employee connects their own Google account — Drive, Gmail and Google Analytics — through the organisation's own OAuth client. A connection flow started in one session cannot be completed by another.

There are no Microsoft, Teams, WhatsApp or Telegram tiles here. Those ingests exist, but they are not wired to the corporate surface.

The boundaries, stated up front. Drive is read five years back, up to 10,000 files per sync, with a 5 MB per-file download cap. Images, video, audio, archives, Forms and Sites are skipped outright, and there is no OCR — a scanned, image-only PDF genuinely extracts to nothing. Gmail defaults to the last 90 days, with spam and trash excluded. Extracted text is written to a path containing that employee's own address; that is the filesystem-level isolation layer, sitting underneath the database one.

Step two · learn

It builds a memory

Three stages, all offline — they run ahead of the question rather than during it.

Extraction to text

Google Docs export as markdown. Sheets export as XLSX first, so every tab survives rather than only the visible one, and are then flattened. Slides, Drawings and Apps Script are read, along with uploaded PDF, DOCX, XLSX and PPTX.

The input is untrusted by definition, so the Office parsers refuse a decompression bomb and refuse any embedded DTD or entity, and PDFs are bounded at 5,000 pages. When an extraction fails it raises and is recorded as a skip reason — a failure is never written to disk as though it were the document's own text.

Notes, written by the model

A background builder reads each file's extracted text, sends it to Claude, and writes markdown notes into that person's vault. Up to 20,000 characters of a file are sent for summarising, and anything under 600 characters is skipped as noise rather than turned into a note that says nothing.

A per-person digest

A map-reduce pass turns the vault into one structured digest per person: each source is summarised on its own, then those summaries are reduced into a single document — a cheaper model on the map step, the strongest one on the reduce. That digest is what gets pasted into the system prompt, verbatim, at the start of every conversation.

Say it plainly — the digest is batch

The synthesis job runs on a 20-minute cycle, but it is a sweep, not a per-person rebuild: each run picks up digests that are missing or stale and rebuilds only a couple of people. On our own deployment the staleness threshold is set to 48 hours, so the summarised brain is realistically hours behind, not minutes.
The Drive and Gmail syncs are themselves a background loop with a one-hour minimum age, so a document saved a few minutes ago may not be indexed yet.
The tools see newer data than the digest does. Filename search, content search and opening a file read the index, so they are as current as the last sync rather than as current as this second. A live Gmail or Drive search goes out to Google and reads what is there now. Either way, if you need today's answer the assistant has to go and look — and it is instructed to do exactly that rather than answer from the summary.

Step three · ask

Ask it

Chat, in the browser, grounded on the memory above.

The prompt

What the model is actually given

The persona, then that person's vault text, then any standing instructions they have saved, then a block describing the assistant's own capabilities and tools — and then the last 20 turns of the conversation.

The loop

Bounded at 8 steps

A turn can search, read, search again, then answer. The tool loop is capped at 8 iterations per turn, so one question cannot become an unbounded crawl. Chat is capped at 500 messages per person per day.

The answer

It names the file

The assistant is instructed to quote and cite the note or file path it answered from, and to say plainly when something is not in the vault rather than fill the gap. That is a rule in the prompt, not a mechanism that enforces itself — and we have not measured how often it holds, so we are not going to quote you a number for it.

On the model. Every turn goes to Claude through a proxy running on the same machine, on Sumit's Claude Max subscription. We deliberately do not print a model tier on a sales page, because the tier is a configuration value that changes; the honest statement is Claude, via that proxy. If the proxy is unavailable a degraded turn can fall through to a different provider — Security sets out exactly who sees what.

The toolset

What an employee actually gets

This is the complete list for a corporate account — not a representative sample, but the whole set as it is defined in the code.

search_files

Find a document by filename across everything indexed from Drive and Gmail.

search_content

Keyword search inside the extracted text of files, returning a snippet around the match.

read_file

Open one file and read it, up to 50,000 characters.

web_read

Read a public web page, an X post, a Reddit thread, YouTube metadata or a GitHub repository. There is no browser on the server, so a page that renders itself with JavaScript comes back empty.

list_recent_emails

Work through recent mail from the index, filtered by sender, subject or age.

gmail_search

Search their own mailbox live, in Gmail's own query syntax.

gmail_draft

Write a draft into their own Gmail and return its id. It does not send.

gmail_send

Send a draft they confirmed. It takes a draft id, and nothing else.

gmail_organize

Apply and remove labels — archive, mark read, star, file under a name.

drive_search

Search their own Drive live, by name or content.

save_preference list_preferences delete_preference

Standing instructions a person teaches it once, re-injected into every later conversation.

mark_chat_resolved

Close out a thread of work.

And what is deliberately not in it. The employee toolset excludes the tools an owner gets: it cannot fire an agent from chat, and it has no outbound campaign tools at all. Agents are run from the dashboard by someone who was granted them, not by asking the assistant nicely.

Acting, not just answering

It acts in your mailbox, with a human in front

We are not going to tell you it never sends email. It can. Here is the exact shape of the constraint instead.

Draft

It writes, into your own Gmail

The assistant composes a reply as a real Gmail draft in the employee's own mailbox and gets back a draft id. Nothing has been sent to anyone at this point.

Send

Only a draft you confirmed

The send tool accepts a draft id and nothing else — it structurally cannot send free-composed content, only a draft that already exists. It is instructed to show you the exact recipient, subject and body and to get an explicit yes first.

Organise

Labels yes, destruction no

It can archive, mark read, star and file mail under a label. Trash and Spam are refused — the tool organises with labels and cannot delete or quarantine anything.

The accurate one-line version, and the one we would want a buyer to repeat back to us: it drafts, you approve, and the only thing it can send is the draft you approved.

The recurring half

Agents do the work that repeats

A chat window is the wrong place for work that happens every morning. Globus ships a fleet of agents an administrator can grant to one person, a team or everyone — sales-desk briefs, CRM audits, editorial QA, content and video drafting — and an on-demand run is forced into draft mode, so it writes to a review queue instead of posting anywhere live. Be clear about what they are, though: these are the agents built for Sumit's own companies, made grantable to staff. This is not yet a builder you point at your own processes. Which of them are granted today, which are genuinely producing, and which are still an empty console is set out in full → /agents

The part most vendors skip

How retrieval actually works

If you are evaluating this against something with "RAG" on the tin, read this section carefully. We are not that, and we would rather you found out here than in month two.

Filename matching

A database LIKE match on filenames. The query is split into up to four tokens, files matching any token come back, and results are ranked by how many tokens hit.

A literal substring scan over recent files

Content search opens at most the 500 most recently modified extracted files — spreadsheets, documents and decks first — reads up to 200,000 characters from each, and looks for your search string as a lowercase substring. Where it finds one, you get a snippet around it.

The digest, pasted in whole

On top of both, the pre-built digest is placed into the system prompt verbatim, which is why the assistant can answer a broad question about the business without searching at all. It is generously capped, and it is the batch artifact described above.

What that means, and what it costs you

This is keyword retrieval. It is not semantic search and it is not RAG. There are no embeddings, no vector database and no similarity scoring anywhere in the system.
A synonym can miss. Searching for "revenue" will not find a sheet that only ever says "turnover". The assistant is told to try a couple more phrasings and then to stop and say it could not find the answer rather than keep guessing — but the underlying match is literal.
Content search reaches the most recent slice of the vault, not all of it. A file outside those 500 newest is fully ingested and still findable by name, but that particular scan will not see inside it.
This is a recorded decision rather than an oversight — our own architecture note on it is titled "no embeddings yet". It is the honest current state, and it is the first thing we would change for a corpus that is large and whose vocabulary is inconsistent.

Voice

Real — and not on this portal yet

● Individual product

Built, and the brain is ours

On the individual product at /members/globus the voice path is built and configured. Speech-to-text, text-to-speech and interruption handling are done by a third party, but the thinking is not outsourced with them: every turn routes back to our own endpoint, which loads that person's digest and runs the same tools as text chat.

Replies are deliberately capped short there, because they are being spoken aloud rather than read. One honest caveat: it is not currently carrying traffic — no voice turn has gone through it in the last month — so treat it as built rather than battle-tested, and ask us to place a live call in front of you before you weigh it.

○ Corporate portal

Text only, today

The corporate portal has no voice at all. An employee signing in here gets text chat, and nothing on this page should suggest otherwise.

It also means voice adds a third-party dependency on top of the model provider, which the rest of the product does not carry. We would rather say that than put a microphone on a corporate pitch and let you assume.

Limits

What it is not, yet

A short list, plainly put, so nobody discovers these after signing.

Not built today

No mobile app. There is no iOS or Android application. It is a browser.
No Slack app. The assistant is not reachable from Slack, or from any other chat platform, on the corporate portal.
No browser extension for corporate. An extension exists on the individual product for capturing WhatsApp Web; there is no corporate equivalent.
Teams is ingested but unanswerable. On the individual product, Microsoft Teams group chats can be pulled in and counted — but there is no Teams search tool and Teams messages do not reach the digest, so the assistant cannot answer from them. The data is collected; the model never sees it.
No per-tenant persona. The assistant's character and standing instructions are a single shared definition, written for individual members. It is not yet re-voiced per organisation, and an employee will occasionally notice that.

Connect once. It learns. Then it does the work.