AI and MCP
Every Hex Is an MCP Server.
HexDB ships with its own Model Context Protocol server. Point Claude Code, Claude Desktop or any MCP client at a hex and your agent can explore and query the lattice: what's in it, what it looks like, and the answers to questions about it. No plugin, no proxy, nothing extra to run.
Listen to the podcast - 4:55 min
The Oracle: Keeping AI Latency Off Your Data
Why AI questions are slow and bursty, and what happens when they share a machine with your data. Give one hex the Oracle role and it answers MCP, the chat and A2A, reading what each question needs from the other hexes as the user who asked, while keeping no data of its own. If it's down, the other hexes answer as before.
Connect an Agent in One Line
Make an API key for a user whose roles allow what the agent should see, then add the hex to Claude Code. Any hex in the lattice answers.
claude mcp add --transport http hexdb \
https://hexdb.example.com/mcp \
--header "Authorization: Bearer hxk_..."
# Then ask Claude about your data:
# "Which customers ordered more than three times last month?"
# "What's in the media tessellation, and how is it indexed?"Nine Tools, All Read-Only
And a Chat of Its Own
Turn on the AI chat and the admin UI gets an Ask page: questions in plain words, answered by your configured model with the same read-only tools and the asker's permissions. It has its own spending limits (a cheaper model, tokens per answer and per question, per user and per visitor per day, questions per minute) under the lattice's monthly budget, so a public deployment can't run up your bill.
Listen to the podcast - 5:41 min
AI Security Reviews: Intrusion Detection on a Schedule
Password guessing, logins tried from a dozen addresses, role changes at 3 a.m.: the lattice's AI reads a summary of the audit trail every 1 to 24 hours, never your documents, and tells administrators what looks like an attack, with the evidence and what to do.
An AI that reads your audit trail for attacks.
Password guessing from one address, a login tried from a dozen, a user refused again and again, a sign-in from somewhere new followed by role changes at 3 a.m.: the signs are in the audit trail, if someone reads it. HexDB has the lattice's AI read it for you on a schedule, and tells your administrators what it finds, with the evidence and what to do.
- As often as you like. Every hour, every day, or anything between: pick 1 to 24 hours on the admin UI's Security Reviews page. Each review covers the time since the last, and Review now runs one on the spot.
- It never sees your data. HexDB sums up the window itself (failed sign-ins by address and login, refusals by user, new addresses, activity by hour, administrative changes) and the model reads that summary: counts, logins and addresses, never documents.
- Findings reach you where you work. Anything of medium severity or worse (or the level you pick) is emailed to the administrators over your own SMTP server and sent to the channels you set up: a PagerDuty incident, an SMS through Twilio, a Slack, Teams or Discord message, or a signed webhook. Every review, quiet or not, is kept for 90 days with its evidence, and audited.
- The Oracle keeps watch. With an Oracle in the lattice, the reviews run there, off the hexes that hold your data, within the lattice's monthly AI budget.
Agent to Agent (A2A)
Other agents can ask too, over A2A (Agent2Agent). Where MCP hands an agent the tools, A2A hands it an answer: the lattice's own chat does the lookups and replies.
- A Signed Agent Card
/.well-known/agent-card.jsonnames the skill, the JSON-RPC endpoint at/a2aand how to sign in, signed with the lattice's key so callers can check where it came from. It says nothing while A2A is off. - The Caller's Permissions
Every request signs in with an API key, a session or an OAuth access token, and the question runs as that user: row filters, field masks and the chat's spending limits all apply. Each answer is audited.
- Answers, Streamed or Pushed
SendMessagereturns a finished task;SendStreamingMessagestreams each lookup and the answer as it's worked out. Or leave up to ten webhooks and the task is posted to them when it's done, to public HTTPS addresses only unless you allow private ones. - A2A 1.0 and 0.3
Both protocol versions work, so agents built on either SDK can ask. It stays off until an administrator turns on the chat and
a2a.enabled.
# The agent card: what this lattice answers, and how to ask
curl https://hexdb.example.com/.well-known/agent-card.json
# A question, answered as the API key's user
curl https://hexdb.example.com/a2a \
-H "Authorization: Bearer hxk_..." -H "A2A-Version: 1.0" \
-H "Content-Type: application/json" \
-d '{"jsonrpc": "2.0", "id": 1, "method": "SendMessage",
"params": {"message": {"messageId": "1", "role": "ROLE_USER",
"parts": [{"text": "How many orders shipped last week?"}]}}}'