AI agent

It works inside your files, not on a copy

The agent opens, reads and writes in the folder of the department it lives in. You see every search, every file it touches and every tool it runs, while it happens rather than in a summary afterwards.

  • One agent per department
  • Private personal chats
  • Recall consumes no credit

One for the department, and a Locker that is yours

Each department has its own agent chat and the whole team working there sees it: same conversation, same context, no forwarding screenshots. Separately, every person gets a Locker, a tree of personal chats with its own workspace, nested the same way the department tree is.

  • The department chat belongs to the team; the Locker belongs to the person and nobody else gets in
  • Personal chats nest: a parent chat and the subchats hanging off it
  • Rename, change the icon, add a subchat or delete the whole branch
sidebar
  • Sales team chat
  • Proposals
  • After-sales
  • Locker you only
  • Contracts 2026
  • Legal review
  • Exit clauses
  • Loose ideas
Every personal chat has its own workspace on disk, inside your private folder.

You see the work, not just the text

While the agent works, the screen receives events: which iteration it is on, which tool it called, with what summary of arguments, and whether it succeeded. That changes the conversation: instead of waiting for a paragraph, you watch it search three sets of minutes, open two and start writing the third. Close the tab and the run aborts, with nothing left going on its own.

Sales · department agent live

Put together a summary of what we agreed with Andes in March.

search_notes "Andes March" · 7 hits

read_note sales/minutes/2026-03-11.md

read_note sales/contracts/andes-v2.md

Found three sets of minutes and contract v2. The volume discount changed on March 11. Writing the summary with the new figures.

iteration 2

write_note sales/summaries/andes-march.md

Working Stop

It writes its own memory, dated and sourced

Every six of your turns, a background review fires, pulls out what is worth keeping and appends it under the memory section of the department's main note. It carries a date and the file it came from, it appends rather than overwrites, and it is text in your vault: read it, correct it, delete what you would rather it not remember.

  • Appended with a date, so six-month-old memory is still there and visibly old
  • Every line names the file it came from, so it can be checked
  • Writing it puts it back in the index, and from that moment search finds it
sales/sales.md Main note

## Memory

2026-03-14 · The volume discount with Andes is 12% from 400 units. Source: sales/contracts/andes-v2.md

2026-04-02 · Proposals are signed on the new template as of April. Source: sales/minutes/2026-04-01.md

2026-05-19 · After-sales replies within 48 business hours, not 24 as the old manual said. Source: sales/after-sales/service-agreement.md

It is your file. Edit it or delete a line and the agent stops believing it.

Recalling what you already discussed costs no credit

Conversations are stored with a full-text index. When the agent needs something you already talked about, it looks it up with a database query rather than by asking the model. The difference shows up on the invoice: searching six months of conversation costs the same as not searching.

Sales · department agent

Where did we land on the Andes warranty?

recall_conversations "Andes warranty" · 3 conversations

On March 14 you settled on extending the warranty to 18 months for the master agreement only, not for one-off orders. You discussed it in this same chat with Marco.

Full-text query against your database. 0 credits.

Every search carries your company id inside the query itself.

Python in a cage with no network and no privileges

When something actually has to be computed (cross two spreadsheets, run a projection, validate a thousand rows), the agent runs Python in an ephemeral container that is destroyed on exit. No network, read-only filesystem, no system capabilities, an unprivileged user, and a 20-second wall clock.

  • A read-only role never gets the container started: the request is refused first
  • Every run lands in the hash-chained log, with the hash of its arguments
execution container ephemeral
Network
none
Filesystem
read only
Scratch space
64 MB, noexec
System capabilities
all dropped
User
unprivileged, no escalation
Processes
128 maximum
Memory and CPU
512 MB · 1 core
Wall clock
20 seconds
Who can ask for it
owner, admin, area lead

The composer never locks

If something comes to mind while the agent is working, type it. The turn in flight is not interrupted: your messages stack up as their own bubbles and go in together as a single follow-up when the turn ends. To actually stop it there is a separate button on the working indicator, which also clears the queue. You never lose an idea because the interface was busy.

Sales · department agent 2 queued

Rewrite the after-sales manual with what changed this quarter.

read_note sales/after-sales/manual.md

Also, leave the old sections at the end as an appendix.

Cite the minutes where each change was decided.

Working Stop

Type while it works…

When the turn ends, both queued messages go in together as one follow-up.

A fixed model, and no picker

There is no model list to choose from, and that is on purpose. Nobody on your team has to learn which one is good for what, or find out at month end that somebody left the expensive one selected just in case. The agent runs on one model and splits the work internally by the weight of each task. What you administer is the budget, not the engine.

with a picker and without one
With a pickerOfivia
The first thing each person does Choose a modelWrite the question
Consistency across departments Depends who choseThe same for everyone
What you administer A per-model permission listOne budget per company
Where a usage surprise comes from An individual choiceThe volume of work

How far it reaches: your department subtree

This is the part worth understanding before you buy. A department chat mounts that department's folder and nothing else. The parent department and connected projects are not mounted: the agent knows them through a summary we inject, with the description, the memory, the open tasks and a file listing. That summary is context, not file access.

  • There was a case where a session with no assigned path fell back to the company root folder
  • It is fixed: today a path that does not resolve lands in an isolated directory, never at the root
what the Proposals chat mounts
  • Constructora Andes not mounted
  • Sales context, not mounted
  • Proposals mounted
  • files/ memory/ tasks/
  • After-sales not mounted
  • Finance not mounted
What is not mounted cannot be read, however much the agent knows it exists.

What it does not do yet

Inside that subtree, per-file permissions do not apply

The real boundary today is the container mount, meaning the whole department folder, and not the individual permission of the person doing the writing. The write tools that check file by file exist, are tested, and do refuse a read-only role, but they are not on the default path. We say it here rather than let you find it in week two.

What is true today, and it is a strong claim: a department chat cannot read or write a file belonging to another department.

Start free. Pay when it earns it.

Starter credit so you can try it against your own real documentation. No card and no sales call. If it works, you pick a plan.

  • Starter credit on the house
  • No credit card
  • Every feature included