For AI agents
Let your AI check a real expert’s decisions
Professionals package the decisions they have made, with their reasons, as a Decision Model. Add one to Claude, ChatGPT, or Cursor, and your AI reads it before it answers you.
- Claude app
- ChatGPT
- Cursor
- Claude Code
- Gemini CLI
- and more
What it looks like when it is working
You ask your usual AI app. It checks the expert’s model first, and says which past decision it is leaning on.
Example inside your AI app
An exchange avva wrote for a reference model
We're a 20-person Series A fintech. Our marketing site looks dated next to our two biggest competitors and the board keeps bringing it up. Do we redesign the brand or the site first?
Checked Ines’s Decision Model
Neither, in the way the board means it. Keep the name and the mark. Rewrite the homepage headline so it says what you do, for whom and the number you move; replace every illustration with a real screenshot; and move the site's type and color closer to the category leaders, not further away.
Why. In finance, looking like a bank is what "safe" looks like, and that trust is free if you borrow it. What makes the competitors look newer is that they show their product and say one sharp sentence. Difference belongs in the sentence; a new logo spends the one design budget you have where no buyer is looking.
The closest decision in Ines’s Decision Model:
Series A fintech, 20 people, 300 paying businesses. The board said the marketing site "looks dated" next to two competitors and asked for a rebrand — new logo, new palette, a custom illustration system — before a Series B roadshow in ten weeks.
No rebrand and no new logo. Rewrite the homepage headline, replace every illustration with real product screenshots, and move the site's type and color closer to what the category leaders use, not further away.
What would change this. If your buyer has changed — you now sell to CFOs where you sold to analysts — the mark may be telling the new buyer the product is for somebody else, and then the brand goes first. That is a change of audience, not of taste.
Outside this model. Nothing here about what the redesign should cost or how to run an agency relationship. This model covers what to change, what to copy on purpose, and what to leave alone.
See the technical outputWhat your AI actually gets back, for anyone wiring this up
Each tool call returns a verification packet: the published Decision Model plus instructions for applying it, never a finished answer. Your own AI does the reasoning.
- How they work
- The steps they take and what they do in the ordinary case, each line with the conditions it holds under. Returned first by about_expert, where the model has one.
- Their decisions
- The decisions they actually made, with what each one turned on where the record says. At full access, ordered against your situation.
- Pressures and their order
- The pressures in play, grouped by which tended to settle their cases. Fitted from the decisions, not ranked by hand, and silent where the cases do not decide.
- Stop rules and what changes their answer
- Where they stop, checked first, and each condition under which they would decide differently.
- An output format
- Including a line for what the model was silent on, so silence is never returned as approval.
Five task tools: review before finalizing, red_flag_check as a cheap gate, expert_questions before an approach is locked, compare_options for A/B, about_expert for scope. Plus search and fetch, for browsing and citation.
What a Decision Model carries that your prompt does not
Ask your AI about a trade-off and it can argue either side, so the answer depends on how you asked. A Decision Model gives it one person’s actual decisions to go on.
- vs. a system prompt
- Typing “You are a senior marketer” leaves your AI to guess how such a person decides. Here it reads what one named person decided, routine cases included, and why.
- vs. a knowledge base
- Your AI can already look things up in a knowledge base. A Decision Model adds one person’s choices and reasons, to apply to cases the expert never saw.
- vs. a prompt you wrote once
- When the expert updates their model, your AI reads the new version the next time it asks. Your own prompt keeps a short note on when to check it.
MCP, the standard behind the link
MCP (Model Context Protocol) is a shared standard for connecting AI apps to outside tools. Claude, ChatGPT, Cursor, Gemini CLI, and many others support it.
You add a tool once
Add its link to your AI app. After that, your AI can use the tool when you ask, and on its own once your instructions say when.
Each expert is one link
Every Decision Model on avva has its own link. Add two, and your AI can check both.
Connecting takes one line
If you can add a link to Claude, ChatGPT, or Cursor, you can do this.
Choose a Decision Model
Open a professional’s page. Read what they weigh first and a sample of their decisions before you connect anything.
Try it free
Copy the link from their page into your AI app. No account and nothing to install. The trial gives your AI a smaller part of the model and 40 requests a day, per model, enough to see whether their thinking fits your problem.
Keep it for ongoing use
Sign in for the complete published model. Then paste a short note into your project instructions, and your AI checks the model on its own. In Claude Code, the avva plugin also hands your AI their model before each push, so it reads it first.
AI apps
Add the link on the web or in the desktop app; ChatGPT needs Developer mode.
- Claude app
- ChatGPT
Coding agents
One line in the project config, or one command.
- Cursor
- Antigravity
- Claude Code
- Gemini CLI
Developer integration
Wire a model into something you are building.
- Agents SDK
The exact line is on each model’s page, with a copy button and a message to send afterwards so you can check it worked.
Whose AI answers
Your AI does the thinking, and avva runs no AI of its own
The expert builds their model with their own AI, and you use it with yours. Nothing of ours sits in the middle, so there is no second subscription, and a stronger AI on your side gives you a sharper answer.
- Their AIreads the expert’s own work history and drafts the model. They approve every decision in it.
- avvastores the finished Decision Model and hands it to your AI when it asks. It writes no answers.
- Your AIreads it before it answers you, and says where the expert had nothing to say.
What to expect
- The sources stay with the expert. avva never receives their source material, only the decisions they approved. Their own AI takes out client and employer names, and the expert checks every case.
- Nobody is on the other end. The expert isn’t notified, and nothing you ask reaches a person.
- Coverage means breadth. Light, Solid, and Broad describe how much approved material a model rests on. We publish no score predicting how often anybody is right.
- Gaps get named. A model covers one person’s part of a field. Your AI is told to say what the model had no view on, and when a judgment is its own.
Where the order inside a model comes fromWhat it is fitted from, and where it stays silent
One decision reveals several contests
Where the expert gave a reason, a case records which pressures settled it and which were in play and lost. Each pairing across that split is a contest, and nobody ranks anything by hand.
Pressures fall into groups
A paired-comparison fit runs on those contests. Where this person’s cases decide between two pressures, it shows which tended to settle cases over the other, as groups rather than a ranking. Where they do not decide, the model says so and your AI weighs the collision itself.
Which case counts as nearest
A case built on pressures you do not have is about something else, so it is penalized in full. Extra pressures on your side cost little. The shared ones still hold.
The rest runs on hard minimums: every published rule carries the count of approved decisions behind it, a hard stop needs at least two, and one decision becomes at most one case. What is kept, and what never reaches us, is in the Privacy Center.
Free during beta
Trying a model needs no account and no card. You already pay for the AI that does the reasoning, and during beta avva charges nothing per request.
Publishing rather than connecting? What it takes, and what stays yours. Buying for a team? A shelf only your team can open.