Policy
Cedar-based rules decide every agent action before it runs: amount caps, per-action allowlists, and segregation of duties — the agent that drafts an invoice can never be the one that approves it. No decision means deny, by construction.
MCP-native governance for agent actions
Agent Sentry — the governance gate for AI agents on your books
Agents act. Policy decides. Humans approve anything over $10k. Every step is evidenced in a tamper-evident ledger — so your AI can run accounts receivable and payable on QuickBooks without anyone losing sleep.
Evidence
A tamper-evident ledger records intent, decision, and outcome for every action, anchored in an immudb transparency log and sealed to KMS-signed WORM storage. If the evidence can’t be written, the action doesn’t execute.
Approval
High-risk actions pause for a human. Approvers get a Slack deep link, see the exact transaction, and sign off with a fresh-MFA login — an agent credential can never satisfy an approval.
How it works
Your agent
claude.ai / Claude Desktop
connects over MCP and proposes actions — create invoice, record payment, pay a bill.
Agent Sentry
The MCP gate
evaluates policy, writes evidence, and routes anything risky to a human approver.
Your books
QuickBooks
only ever sees actions that passed policy — each one traceable to its evidence record.
Put a sentry between your agents and your ledger
Sign in with your Google Workspace account and connect QuickBooks in minutes.