Route risky
actions before they
run.
Safe Mode gives AI coding agents a controlled path for file writes, command-like tools, Git mutations, package changes, and deploy-like actions when those actions are routed through AgentVeil.
Configured routes only. Not host-wide sandboxing. No claim of complete agent safety.
Don't just stop the agent. Give it the next controlled move.
AgentVeil turns a risky routed request into a bounded correction path, then records what actually happened.
AGENTS.md is not a control boundary. It tells the agent what it should do. AgentVeil applies authority when the action reaches a configured route.
Click a risky action. Watch the boundary answer.
This is the product idea in one interaction: the agent proposes an action; AgentVeil gives the next safe move on the routed path.
Open a pull request instead.
Project rules can say “do not push directly to main.” On a configured Git route, Safe Mode redirects that request to a pull request before upstream changes.
The first useful surfaces for coding agents.
Keep the copy honest: these are controlled when they are routed through AgentVeil, not when they happen through raw shell or unmanaged APIs.
Filesystem
Gate routed writes, deletes, sensitive paths, instruction-file changes, and write-then-execute patterns.
write_file · delete · chmodCommand-like tools
Require approval for routed command tools and block dangerous patterns before execution.
rm · curl | sh · chmod +xGit
Gate routed Git mutations such as add, commit, reset, clean, rebase, and push on a configured repository.
git push · reset · cleanPackages
Require approval for routed package-manager mutations such as install, update, uninstall, and install-script execution.
npm install · pip installRelease / CI
Review routed release, workflow, package-publish, and deploy-like actions exposed through a configured path.
workflow · release · deployMCP and connectors
Bring supported MCP and connector actions through one policy boundary where connected.
MCP · IDE route · API gateShow what is controlled, missing, or bypassable.
Project instructions express intent. Doctor shows whether authority actually exists on the configured route, and exposes raw shell or direct access as bypass risk.
Add the controlled route to the terminal you already use.
Choose a supported client, connect one project, then use Doctor to confirm what is routed before the agent acts.
Delegate actions without handing over control.
Start locally. Add hosted records and team-scoped controls when agents move closer to code, credentials, and sensitive systems.
Core
Open-source control layer for AI-agent workflows.
- Open-source AgentVeil CLI
- Local approval loop
- Basic redirect guidance
- Local evidence and status
- Public fallback path
- No license required
Builder
For solo builders controlling AI-agent development workflows.
- Private policy package
- Advanced redirect controls
- Stronger policy guardrails
- Usage visibility
- 25,000 controlled actions / month
- 14-day trial
Team
For teams controlling agent actions across shared development workflows.
- Shared workspace and team policy controls
- Team approval routing
- Shared action history and decision records
- 90-day shared record retention
- 300,000 routed action decisions / month
- Development workflow controls
Enterprise
For organizations with custom security, deployment, and compliance requirements.
- Custom controlled-action volume
- SSO / SCIM options
- Custom retention and legal hold
- Security review
- Private deployment options
- Custom support and SLA terms
Monthly allowances are based on routed action checks: agent actions evaluated by AgentVeil before execution. Routine checks can allow actions automatically; higher-risk checks can require approval, block, or redirect. Local-only Core usage is free.