# avva > Decision Models built from what one expert actually decided — the cases, the forces that settled them, and the rules behind them. An agent consults one before it finalizes a decision in that domain. avva does not run a server-side LLM; the calling agent does the reasoning. avva is a marketplace: one Decision Model equals one public expert page. Buyers try a model by connecting its MCP URL with no key — it answers at once with a slice of the model — from Claude, ChatGPT, Cursor, Antigravity, or another MCP client, then sign in for the complete published model. Experts extract judgment from approved past decisions through Studio; published models are sterile (no source quotes or private cases). Coverage badges describe breadth of evidence, not predictive accuracy. A Decision Model replaces the "You are a senior who…" system prompt a user would otherwise write — not the person: the calling agent still reasons as itself and never answers as the expert. It carries one professional's everyday calls and their departures from the usual answer, what they weigh first, where they stop, what flips a call and what they ask before deciding. What stays in the user's own instructions is a short note on when to consult it. A model changes only when its expert republishes it: the URL stays the same and serves the current version on every call, so a connected agent reads an update the next time it asks, with nothing to reinstall, and each packet says when the model was built. Nothing notifies the user that it changed. ## Product - [Home](https://avva.chat/): What avva is, why a prompt is not enough, and how it works in three steps. - [For AI agents](https://avva.chat/for-ai-agents): When to connect a Decision Model, how it differs from the system prompt a user would otherwise write, what a connected answer looks like, and how to set one up. - [For experts](https://avva.chat/for-experts): What publishing takes, what leaves the expert's machine, and what stays theirs. - [For teams](https://avva.chat/for-teams): A private shelf of Decision Models for one company, readable by invited members only. - [Expert marketplace](https://avva.chat/experts): Browse published Decision Models and open profiles. - Catalog MCP server (discovery only): `https://avva.chat/mcp-catalog` — tools `search` and `fetch` return experts with their profile and per-expert connector URLs. Consulting runs on the expert's own connector, not here. ## Models - Kirill Dobrinov — B2B Product Management: `https://avva.chat/mcp/b2b-product-management-kirill-dobrinov` ([page](https://avva.chat/experts/b2b-product-management-kirill-dobrinov)) ## Trust and legal - [Privacy Center](https://avva.chat/privacy): What transits Studio, what is stored, sterile publication. - [Terms of Use](https://avva.chat/terms): Accounts, published models, disclaimers. - [Help](https://avva.chat/help): Plain answers on what a Decision Model is, trial and keyed access, what publishing takes and keeps, workspaces, and who runs avva. ## For agents - MCP runtime: `https://avva.chat/mcp/` — packets carrying how the expert takes up a decision (where their model records it), their past decisions and what settled each, then their stop rules and criteria, for your agent to reason with rather than a finished answer (expert_questions, red_flag_check, review, compare_options, about_expert), plus `search` and `fetch` for browsing and citation. Without a key a response is a slice: the rules and, from `review` and `compare_options`, five cases picked by word overlap and a one-line index of every case: where it holds situations closer to yours, call that tool again with `cases` set to up to five of their ids to read those in full. With a key from a signed-in avva account it is the complete published model. - When to consult a connected model: Consult it CONTINUOUSLY in this domain, once per decision, not once per message; skip trivially reversible work (renames, typo fixes, formatting); when unsure, err toward calling. START of a task — `expert_questions`, before locking an approach. WHILE working — `red_flag_check` on drafts, plans and diffs. BEFORE finalizing or shipping anything significant — `review` with context + proposal; your verdict comes first, then your delivery. A/B choices — `compare_options` (a stop rule of BLOCK severity disqualifies an option before the criteria are compared). Scope — `about_expert`. - Limits and keys: without a key, 40 requests a day, per model; with a key, 1,000 requests a day, per model. A key is issued after your user signs in on the expert's page (`https://avva.chat/experts/`) and lasts 90 days; a key a workspace member issues for an organization's model lasts 365 days. To use it, pass it as the `Authorization: Bearer ` header, or — where a connector dialog has no header field (ChatGPT), or the URL is simply the easier line (the Claude app) — on the end of the URL: `https://avva.chat/mcp//`. - In Codex: add the server as a `[mcp_servers.]` table with its `url` in `~/.codex/config.toml`, and make the file's first line `tool_output_token_limit = 120000`. Without it Codex keeps about 12,000 characters of a result and drops the middle, which on a packet is most of the model. - Server card: `https://avva.chat/.well-known/mcp.json` — the three servers, their tools and prompts, limits and where a key goes, as JSON. - Studio extractor: `https://avva.chat/mcp-studio` — expert-side build path (not for buyers). - Sitemap: [https://avva.chat/sitemap.xml](https://avva.chat/sitemap.xml) ## Optional - [Studio UI](https://avva.chat/studio): Browser wizard companion to /mcp-studio (noindex).