Every dangerous action your AI agent attempts requires your explicit approval via push notification. Default-deny. No SDK required. The agent cannot bypass the gate.
See How It WorksMost guardrails operate at Level 1 or 2. Agents eventually ignore prompt rules and skip voluntary gates. Agent Vigil operates at Level 3.
"Please don't delete important files."
Agent ignores eventually.
"Are you sure you want to delete this?"
Agent can skip the gate.
Action is physically blocked until a human approves.
Agent cannot bypass the gate.
One HTTP endpoint. One push notification. One tap. That's the entire integration.
Your pre-action hook sends the action details to Agent Vigil's API. Any language, any framework.
Your phone buzzes with what the agent wants to do. Project name, action type, full context.
Approve or deny with a single tap. No response in 5 minutes? Auto-deny. That's the contract.
The hook unblocks. Exit code 0 (approved) or 1 (denied). The agent proceeds or stops.
THE ENTIRE INTEGRATION
# Your hook sends one HTTP request. That's it. curl -X POST https://api.agentvigil.com/v1/review \ -H "Content-Type: application/json" \ -d '{"key":"av_your_api_key_here","action":"delete_file","context":"/src/components/legacy"}'
The request blocks until you approve, deny, or the 5-minute timeout expires.
When an agent attempts a dangerous action, your phone buzzes with the full details. Approve or deny without opening the app. Get back to what you were doing in under 5 seconds.
Agent Vigil is an infrastructure enforcement layer, not another prompt rule. These are the design decisions that make it work.
Allowlist, not blocklist. Missing a rule means the action is blocked, not exposed. The same philosophy that makes firewalls work.
One HTTP POST is the entire integration. Use curl, Python, Node, Bash, or anything that speaks HTTP. Works with any agent framework.
No response in 5 minutes? Action denied. No human available means no action executed. That's the safety contract.
Action data is transient — held in memory only during the review window. Never stored. Never logged. Never shared with third parties.
Push notification delivered within 2 seconds. Response back to the hook within 2 seconds of your tap. No perceptible delay.
Claude Code, OpenAI Codex, LangChain, custom agents. If it can call a pre-action hook, it works with Agent Vigil.
One plan. Full access. No tiers, no usage limits, no surprises.
Billed monthly via Apple In-App Purchase. Cancel anytime.
Download Agent Vigil, create a project, and run the curl test. See your first push notification in under 10 minutes.
Coming to the App Store