The core MCP server is live
You built and deployed it. AI clients (Claude, Cursor) connect by URL + bearer token and call audited tools that operate the real database.
One contract, every AI client, full audit
The AI never touches the database directly. It speaks MCP to the server, which checks the caller's token + scope, runs a named tool, and writes an audit row — every time.
What you control as super admin
A single screen to govern the whole MCP layer. Below is the live tool registry exactly as the console would render it — toggle a tool, see its scope, risk, approval gate, and call volume.
| Tool | Server | Scope | Risk | Approval | Calls | Enabled |
|---|
The eight controls the console gives you
Issue / revoke a bearer token per principal (Mike, each lane, verifier, a vendor). Rotate on a click.
Enable or disable any single tool instantly — the kill switch for one capability.
Which principal/role may call which tool — the matrix, editable.
Flag high-risk tools (money, deploy) to require your one-tap OK before they execute.
Every call — principal, tool, args hash, result, time — filterable and exportable.
Per-principal daily limits; alert at 90%; auto-pause on breach.
Calls per tool / principal / day; spot anomalies and unused tools.
Disable a whole server in one click if anything looks wrong.
From 7 tools to a full operating surface
Today's 7 cover board / memory / QA. Best-in-class means the AI can safely read and act across the whole platform — each new tool scoped and risk-rated. Live counts below.
Reads are low-risk and ship first; writes that move money or publish are approval-gated and ship behind their own servers (next section).
Split by blast radius, not convenience
One server for everything is a single key to the whole kingdom. Best practice is a server per domain — its own token, scopes, approval policy, and kill switch — so a problem in one can't touch the others.
Five phases, each behind an approval
Nothing ships without your sign-off. Each phase is a clean increment you can approve, see, and stop.
Phase 1 — Core server
pure-mcp with audited tools, bearer auth, MCP handshake. Built & deployed.
shipped & livePhase 2 — Admin console + read tools
Governance screen (toggles, scope matrix, audit viewer, usage, kill switch) + read tools (deal.read, lead.create, doc.status, search.listings). Open the Admin Console
APR-MCP-1 donePhase 3 — pure-comms server
email.send / sms.send / comms.log + comms.notify, audited & rate-aware. Delivers once provider keys are set.
APR-MCP-2 donePhase 4 — pure-money + pure-deploy
release.cut/publish/list + cda.read. publish is approval-gated + critical; queues a deploy ticket until the Netlify token is set.
APR-MCP-3 · needs APR-3 token for real deployPhase 5 — pure-mls + pure-connect
listing.check (RESO), referral.read, alert.read. Read-mostly external surface.
APR-MCP-4 doneIdeas to make the MCP server best-in-class
Beyond the 22 live tools — the upgrades that turn it from "AI can operate PURE" into "PURE runs itself, safely."
Online ticket intake AI auto-build
A public/admin "submit a ticket" box that writes to the board via board.dispatch; the MCP server auto-routes it to the right AI tier (haiku/sonnet/opus) and lane, and a code/design engine picks it up. You type a request, it gets built. This is the big one.
Memory + dictionary tools
memory.get/set are live — add dict.lookup/add and a nightly memory.sweep so the AI keeps its own facts current and the PURE Dictionary self-maintains.
Scheduled / triggered runs
Cron-style jobs: nightly MLS data check, morning lead-routing sweep, weekly partner-referral nudges — each an MCP tool call, logged + reversible.
Approval inbox for gated tools
When a gated tool (e.g. release.publish) is called, it pauses and drops an approval card in your Command Center; one tap runs it. Human-in-the-loop for the risky 5%.
Rollback tool + deploy registry UI
A deploy.rollback tool reading pure_deploy_registry — one call reverts any deploy to its logged prior version/sha. Plus a visual registry timeline.
Rate limits + spend caps enforced
The per-minute + per-day caps in pure_mcp_settings enforced in the guard, with live spend metering per tool/lane so the AI never runs away.
All 5 phases are built and deployed
As of Tue 6/17/26, every approved MCP phase is live — 22 tools across 6 servers, agent-deployed (pure-mcp v1.2, version 8), server-side security (master kill + per-tool enable + scope), all audited. The list below stays for the record.