Today Pure Chat.ai converses, sees images, and drafts tickets. Best-in-class means it can reach everything we've built — the board, the dictionary, 156 memory facts, the live MCP tools, and the open web — reason across them, and take gated action, all from acquisto.biz on our own API.
Pure Chat.ai stops being a sandbox. Every surface we've built becomes a tool it can read — and, with approval, write. Nothing here is hypothetical; these are live tables and functions in our Supabase project today.
Read, draft, dispatch, claim, update — the same board every lane uses.
PURE canon — org model, brand rules, transaction terms. It speaks our language correctly.
Standing facts, doctrines, decisions across every session — so it never forgets how PURE works.
The pure-mcp server: deals, listings, docs, leads, approvals, QA, rate limits — already deployed.
Native server-side web search + fetch — current facts, market data, comps, anything online, with citations.
Reads pasted screenshots & contracts; reaches FUB, Matterport, Maps, Zoho, Analytics via Composio.
Every message runs a reason act observe loop on our own server (the chat Edge Function, holding the Anthropic key). Claude decides which PURE tools to call, calls them, reads the results, and only then answers — grounded in real data, with any write surfaced for your approval.
"What's blocking the Smith closing, and draft the agent a status note." · Paste a screenshot · "Find 3-bed comps near this listing."
Claude plans: look up the deal, check the doc status, search the dictionary for canon terms — then gather what it needs.
Tools run against the live pure-mcp RPCs and Supabase (service-role, every call audited). Results stream back into the model.
A real answer citing live data and the web — plus an approval card for anything that writes (dispatch a ticket, message an agent, save a fact).
Reads run freely so it's fast and useful. Every write is human-gated through an approval card, scoped by who's signed in (agent / broker / super-admin), and written to the audit log.
| Capability | Tools | Access |
|---|---|---|
| Know the build Answer about tickets, status, decisions | board.list · memory.get · dict.lookup · jobs.list | Read · free |
| Know the business Deals, listings, docs, leads, approvals | deal.read · search.listings · doc.status · approvals.list · rate.status | Read · free |
| Search the world Current facts, comps, market, research | web.search · web.fetch | Web · cited |
| Move work Create / dispatch / update tickets & leads | board.dispatch · ticket.update · intake.submit · lead.create | Write · approve |
| Reach out Notify a person, send branded email / SMS | comms.notify · email.send · sms.send · lane.mail | Write · approve |
| Remember & define Save a fact, add a canon term | memory.set · dict.add | Write · approve |
The intelligence is only half of it. The interface makes the agent legible, trustworthy, and fast — so you always see what it touched and stay in control.
Inline step chips — "looked up 3 tickets," "searched the web," "read the Smith deal" — so every answer shows its work.
Any write pauses on a card: the exact action, editable, with Approve / Edit / Cancel. Nothing happens behind your back.
/ticket · /board · /find · /define · /remember · /web — power moves one keystroke away, plus smart suggested follow-ups.
Web answers link their sources; data answers name the record (deal #, ticket id) so claims are checkable.
Tokens stream as Claude thinks; the PureLoader shows during tool calls. It feels alive, never frozen.
It knows who's signed in (Mike, Shana, an agent) and only offers tools that role is allowed — broker vs agent vs super-admin.
We ship value at every phase — nothing waits for the whole thing. Lane A (Design) owns the experience and the live grounding; Code owns the agent loop, tool execution, and gated writes.
The foundation is live as of this session.
chat Edge Function, key held server-side)Before any backend change, make Pure Chat genuinely knowledgeable today by feeding it a curated, live context pack.
Upgrade the chat function to a real tool-use loop over the live read tools, plus native web search.
web_search + web_fetch with citationsLet it act — safely. Writes return a proposed action the front-end renders as an approval card.
Extend reach to outside systems and make recall semantic.
Sent directly to Code via lane-mail. The conversational core is already live — this is the agent layer on top of the existing pure-mcp tools and pure_chat.backend_plan.
chat function — Anthropic tools schema mapped to pure-mcp RPCs; execute observe continue until final answer.web_search + web_fetch server-side, return citations.