Atra Agent's ten “desks” are powered by Anthropic's Claude models. Every response — listing copy, outreach drafts, CMA narratives, checklists, commission summaries — is generated text, not a database lookup or a verified fact. The model can be wrong, can misunderstand your request, and can occasionally state something confidently that isn't accurate. Every desk is instructed not to invent specific facts (prices, comps, dates, names) that you didn't provide, but that instruction is a prompt-level guardrail, not a technical guarantee — you are always the last check before anything goes out.
The Listings/CMA desk does not have a live connection to any MLS or IDX feed. Every comp, price, and market figure it uses comes only from what you type or paste into the conversation. The same is true for Lead Gen and Client Experience with respect to your CRM — nothing is pulled in automatically. If a number looks authoritative, it's because you gave it that number, not because the Service independently verified it.
The Marketing, Content, and Listings/CMA desks run an automated, non-blocking screen for common Fair Housing Act risk language (familial status, race/national origin/religion references, disability-related exclusions, and classic “steering” phrases like “safe neighborhood”) after generating a response. If something is flagged, you'll see a visible warning with the specific phrase and category — the content is never automatically blocked or edited, so you stay in control.
This screen is a keyword/pattern match, not a legal review and not a second AI judgment call. It will miss violations phrased in ways not on its list, and it can flag legitimate language that happens to match a pattern. A “no flag” result does not mean the content is compliant, and you are responsible for Fair Housing compliance regardless of what this screen does or doesn't flag.
The TC Assistant desk is restricted to checklists, timelines, and general explanations of what a standard clause or contingency typically covers. It is instructed to decline requests to draft, finalize, or interpret binding contract language for your specific transaction, and to redirect you to your broker or a licensed attorney instead. Nothing this desk produces is legal advice, and it should never be used in place of official association forms or attorney review.
Any number-bearing output from the Financials or Listings/CMA desks (totals, pricing ranges, summaries) is shown with a persistent on-screen reminder that it's an AI-generated estimate to verify independently. Financials desk output is organizational support only, not accounting, tax, or investment advice — confirm with your accountant before relying on it for filing or financial decisions.
Some desks let you attach an image or PDF (e.g. a listing photo or a contract page) to a message. The file is sent to Anthropic's API for that specific request so the model can read it, and is not stored in our database as raw file data — only a short text placeholder (the file name) is kept in your conversation history. The file is not automatically re-sent on later messages in the same conversation; if you want the model to reference it again, you'll need to attach it again.
If you save an “agent context notes” field in Settings, that text is included in every desk's request to Anthropic on every conversation, so the model has consistent background about you and your business. This is entirely optional and editable at any time.
The Usage tab in Settings shows token counts (exact, from the Anthropic API) and an estimated dollar cost (calculated from hardcoded pricing constants that may not exactly match your actual Anthropic billing — treat it as a ballpark, not an invoice).