For teams
Your senior people can’t be in every review.
One architect, one principal lawyer, one head of brand, and every hard decision waits on them. Put how they decide on a shelf only your team can open, and every colleague’s AI reads it before it answers, without booking an hour of their week.
- Invited addresses onlyexact emails you type, never a whole domain
- Off the public marketplacenot listed, not in the sitemap or search
- Access ends when you remove someonemembership is re-read on every request
What changes for the team
The standard travels
One person’s way of reviewing a contract, scoping a release or pricing a deal stops being something the rest of the team gets by working next to them. The team’s AI reads it instead, on the decisions they are making today.
One colleague’s decisions
A general model answers with the average of the internet. A colleague’s Decision Model carries the decisions that one person actually made, published by them: their practice, not company policy.
It says where they have no position
Outside what the model covers, your team’s AI is told to say so rather than improvise in your senior person’s voice. Its limits are stated, never guessed.
Nobody is on call
No notification reaches the expert, no queue forms, and nothing waits on them being free. Their colleagues’ AI reads the model instead.
How it works
Start a workspace
Name it and you own it. Nothing is shared until you invite someone.
Invite by work address
Exact addresses, up to fifty at a time. No domain wildcards, because inviting everyone at a domain invites contractors, aliases and whoever registers next.
Your experts publish to the shelf
They build a model from decisions they have already made, with their own AI, and choose your workspace instead of the marketplace when they publish.
Colleagues connect their own AI
Each member takes their own connection from the workspace page and adds it to Claude, Cursor, ChatGPT, or whatever they already use.
Who can read it
Each of these is enforced where the request is answered, not by what a page shows.
- The uninvited get the answer a wrong address gets. Not “forbidden”, which would confirm the model is there. A stranger learns nothing, not even that it exists.
- Removing someone ends it on their next request. Membership is re-read every time, so access stops when you remove them rather than when their key expires.
- Each connection is bound to one member. A workspace connection opens that workspace’s models and nothing else; an ordinary personal connection opens none of them.
- avva never receives your source material. Your expert’s own AI reads their work in place. Only decisions they approved are used to fit the model, and what we store is the finished model.
A model outlives the person who published it. It stays on your shelf and keeps answering when they move teams or leave; removing their workspace membership does not take it down. Its author still owns it and can unpublish it or make it public without the workspace owner’s approval. If it leaves your shelf, the workspace owner is emailed. There is no ownership transfer yet, so if that matters to your review, tell us before you start.
Full detail in the Privacy Center. We do not claim perfect automatic anonymization, because nobody can. Review before publishing stays with the expert, and the page says so at the moment they publish.
Two roles: owner and member
What a member cannot see is usually the first thing a security review asks.
Owner
Invites and removes members, sees who is on the list, and sees how often the shelf’s models are used. One owner: the person who created the workspace.
Member
Opens the shelf, takes their own connection, and can use any model on it. They do not see the member list, and they do not see usage.
There is no third tier, no group and no per-model permission yet. Single sign-on, directory sync and exportable audit logs do not exist. If your review needs them, tell us what it needs before you start.
What it costs
Free during beta. Your experts build their models on the AI subscription they already pay for, and avva runs no AI of its own in the middle. Each member’s key allows 1,000 requests a day, per model.
Buying for yourself rather than a team? See how connecting works.