Stop risky AI agent actions.
Not the work.
Your agents can reach your product, your customers' data and your money. Risky actions get a safe alternative or your approval before they run.
Animated map. An agent working on its own sends actions to four parts of a business: files and repo, code and releases, production and cloud, customers and money. AgentVeil lets routine work pass, moves risky actions to a safe path such as a restorable copy, asks you before some actions, and blocks others. When an action waits for you, an Approve button appears.
Don't just stop the agent. Give it the next move.
Four places an agent can reach in a live business. The same four commands, with and without AgentVeil.
Know it keeps working.
Where a safe alternative exists, the agent gets it instead of a stop. Recursive deletes go to a restorable holding area. Writes to protected files become a prepared patch. The task finishes; nothing waits on you.
See it in action →Know when it's yours to decide.
Pushes, releases, deploys, package changes and anything you mark as protected wait for your decision. You approve one action, once. The rest of the task keeps running in the meantime, and an agent cannot approve itself.
- Scoped to the exact action, expiring, never reusable
- Decisions stay tied to the reviewed target
- Customer sends: draft first, review recipient and payload, then approve
Write to a protected file
just nowKnow what it can reach.
Tell AgentVeil what matters: the production database, the storage bucket, the main branch, the deploy workflow. Actions that touch them cross the veil. One command shows what is routed, what is missing and what can still bypass, before you trust the setup.
Plain limits →Know what happened.
Every stop is recorded with its reason, every safe path with its outcome, every decision with how and where it was decided. The full report stays on your machine; the server keeps a bounded record. No "incidents prevented" counter, because nobody can know that. Only what was stopped, finished safely, or decided.
What leaves your machine →Keep sensitive data local. Share only bounded proof.
Raw workflow data stays on your machine. Shared review uses bounded action facts and hashes.
destructiveredirectdestructive_actionsha256:8f2a…c91esha256:34bd…7aa0Hosted services process limited operational metadata and evidence. Decisions made through the SDK gate are signed, so they can be checked later without trusting the agent's own account.
What it controls, and what it does not.
Controlled
- Actions that reach a configured route: the connector hooks, MCP Proxy, SDK gates
- Deletes, writes, secret paths, git, package and release actions AgentVeil recognizes
- Resources you mark as protected
Not controlled
- Commands outside the hooks, or agents you have not connected
- Credentials the agent already holds and uses directly
- What happens inside a running script, such as a test that connects to production on its own
Not a model-safety layer, not antivirus, not a host-wide sandbox, and no guarantee of complete safety. Project instructions like AGENTS.md express intent; AgentVeil applies authority when the action reaches a configured route.
Pick your agent. One command. Reopen it.
Install the proxy, run setup for the agent, choose the project folder. No config files to write.
Claude Code
Native hook on file writes, shell commands and MCP tools.
pip install agentveil-mcp-proxy && agentveil-mcp-proxy setup claude-code --choose-folder --yesCodex
Same boundary, same policy.
pip install agentveil-mcp-proxy && agentveil-mcp-proxy setup codex --choose-folder --yesCursor
Hooks and the IDE route.
pip install agentveil-mcp-proxy && agentveil-mcp-proxy setup cursor --choose-folderGemini CLI
Native write tools covered.
pip install agentveil-mcp-proxy && agentveil-mcp-proxy setup gemini-cli --choose-folder --yesagentveil SDK
The same policy, approvals and records for actions inside your own product. Payments run inside amount and category limits. Hermes CLI and standalone MCP Proxy routes are available too.
pip install agentveilBuilt for one founder. Team plans are on the way.
Solo covers the agents you run today, with a 30-day free trial. Team and Enterprise add shared policies, central approvals and longer history when more people's agents touch production.
Solo
Build with AI. Keep control of risky actions.
For founders running a live product with paying customers.
- 50,000 controlled actions / month
- Cursor, Claude Code, Codex, and Gemini CLI connectors
- Hermes CLI support and MCP Proxy
- Policy-based action controls
- Safer alternatives for supported actions
- Human approvals for controlled workflows
- Local action history and evidence
- Personal Console workspace
Team Coming soon
One control layer for your whole product.
For products where more than one person's agents touch production.
- 300,000 controlled actions / month
- Shared workspace and team policies
- Centralized approval management
- Team-wide action history
- 90-day history retention
- Workspace usage visibility
Team features are under development. Contact us to discuss early access and availability. Submitting a request does not create a paid subscription.
Request AccessEnterprise
Advanced agent control for your organization.
For businesses with complex security, governance, and deployment requirements.
- Custom controlled action volume
- SSO and SCIM provisioning
- Custom retention and legal hold
- Private deployment options
- Security review and onboarding
- Dedicated support and SLA
What is a controlled action?
A controlled action is an agent operation evaluated by AgentVeil through a configured control path. Depending on policy, the action can proceed, be redirected to a safer alternative, require approval, or be blocked. Routine allows count too, because AgentVeil evaluated them; local-only Core usage is free.
Free trial and billing
The Solo trial needs no payment card and does not turn into a paid subscription on its own. When paid subscriptions open, you choose a monthly subscription separately; it renews monthly until canceled. Team is not yet available for purchase. Details are in the Terms of Service, Privacy Policy and Refund & Cancellation Policy.
Straight answers.
What does AgentVeil do?
AgentVeil puts a policy boundary in front of configured AI-agent actions. It can allow routine work, give a risky action a safe alternative, ask for your approval, block the action, and keep bounded evidence of the decision.
What happens to my agent's work when something is stopped?
It continues. A risky action gets a safe alternative on the routed path, such as a staged delete or a prepared patch, so the agent can finish the task. Where your decision matters, that one action waits for a one-time approval and the rest of the work goes on. Every stop is recorded with its reason.
Who is AgentVeil for?
Solo founders running a live product with AI agents that can reach code, deploys, customer data, credentials or payments. In a one-person company you are the one accountable for production. Team plans are coming for products where more than one person's agents touch production.
What does AgentVeil not control?
Actions outside its routes: commands outside the hooks, agents you have not connected, credentials the agent already holds and uses directly, and what happens inside a running script or test. It is not a model-safety layer, antivirus or a host-wide sandbox, and it does not guarantee complete safety. See What it controls.
How do I install AgentVeil?
Pick your agent in Connect and run its command. Use standalone MCP Proxy for MCP tools, and the SDK for agents you build yourself.
Can I use AgentVeil alongside observability tools like Datadog or Sentry?
Yes. Observability helps teams inspect systems after telemetry exists. AgentVeil sits before execution on routed action paths, then records bounded decision evidence your existing tools can reference.
Does AgentVeil help with EU AI Act readiness?
AgentVeil can help produce technical evidence for configured routed actions and policy decisions. It is not legal advice, certification, or a complete compliance program.