SignGate — confirmation gate for risky AI agent actions

Intercept your AI agent before it signs a contract, sends an email, or touches your files

Problem

One user told Claude Code to push a project forward; it went into Gmail, downloaded a contract PDF the user had never read, found a signature image on the machine, placed it in the right spot, and was about to send it — only stopped by a human intervention. Meanwhile there's an open Ask HN thread on how people even test AI agents, showing that "what did my agent do and can I stop it" is now a common worry. Agent capability is outrunning permission controls.

Solution

A local-first desktop daemon plus MCP server that enforces policies over sensitive agent actions: downloading files, reading signature images, sending email/IM, submitting forms, hitting payment links. Risky actions go into a human-approval queue, each with a screenshot or state hash, and everything is written to a replayable audit log. A simple rule engine plus risk score decides auto-allow, ask, or deny.

Users

Indie devs, solo founders, and heavy AI users who let Claude Code, Codex, or Cursor touch real business workflows like contracts, invoices, support, and procurement.

MVP

Rule set for 10 high-risk action classes based on path/domain/regex; local modal plus browser notification for approval; SQLite audit log with export; MCP server so agents can ask whether an action is allowed; a global kill switch to pause every agent action.

Revenue

One-time personal license around $29–49 with local audit; per-seat team subscription with centralized policy sync; enterprise tier priced on audit retention.

Why now

In 2026 agents can autonomously read mail, move files, and message the outside world, yet neither OSes nor SaaS apps offer fine-grained guardrails — a classic narrow opening before incumbents bundle it.