Two products, one rule library
AgentTrail Guard vs AgentTrail OS
AgentTrail Guard protects your machine. AgentTrail OS protects it and keeps the record, free for one developer and ready for a team. Both block the same dangerous actions from the same open rule library. What differs is where the decision is made, what is kept, and what you can show afterwards.
Choosing
Which one is for you?
AgentTrail Guard is the fastest start. AgentTrail OS is free for one developer too, and it is where your agents' history, patterns and proof live.
Choose AgentTrail Guard if…
- You want protection in the next minute, with no sign-up.
- Nothing may leave your machine, whether air-gapped, locked down or on principle.
- You want to read every rule, and every line of the tool that enforces them.
- You are new to agentic coding and want a seatbelt before anything else.
Your agent is stopped before it runs terraform destroy, rm -rf $VAR or reads your .env. Free, local, no account.
Choose AgentTrail OS if…
- You want every agent session on the record and searchable, not a capped log on one laptop.
- You want the product to find the mistakes your agents repeat and propose the rule.
- A second person uses coding agents, and every machine needs the same rules, proven on your own history first.
- Someone asks “how do you govern your AI agents?” and you need an answer with a record behind it.
Start free on your own: the same protection, plus a searchable record and rules grounded in your own history. Add your team when it is time.
Start free with AgentTrail OSAlready using AgentTrail Guard?
Moving from AgentTrail Guard to AgentTrail OS
You don't uninstall anything. AgentTrail OS takes over project by project, and AgentTrail Guard keeps protecting the rest.
- 1
Run setup in the project
npx -y @agenttrail/cli@latest setup --agent claudein each project you want AgentTrail OS to decide in. Setup checks that your AgentTrail Guard can stand down; if it can't, setup stops before writing anything and tells you how to update it. - 2
AgentTrail Guard stands down there, and only there
After AgentTrail OS enforcement is verified, setup records the project on this machine, and AgentTrail Guard steps aside in it. It keeps protecting every other project. If it can't confirm that AgentTrail OS is the one deciding (another server, another credential, the plugin turned off), it keeps enforcing, so you may see two prompts, never none.
- 3
See what your organization doesn't enforce yet
Before AgentTrail Guard stands down, setup lists the rules it enforces that your organization doesn't: your own rules, library rules your organization turned off, and stricter actions. It asks first, and won't stand AgentTrail Guard down at all for an organization with no enabled policies.
- 4
Carry your settings over
Setup writes
guard-import.jsonon your machine and prints its SHA-256. In the app, Governance → Import from Guard reads it in your browser, shows every change in full, and applies only the ones you tick: library rules your organization is missing or turned off, stricter actions as personal copies, and your own rules if you opt in. Changes to rules everyone shares need an Admin.
What changes once AgentTrail OS decides
- When AgentTrail can't be reached
- Your organization's fail mode decides, closed by default: the action is blocked, and your agent says why. AgentTrail Guard lets the action through. Setup and the import screen tell you which way it changes for you.
- An “ask” rule
- Becomes a hold in the approval queue, for up to the rule's timeout (at most 590 seconds). If nobody answers, the decision comes back to you; a headless run is blocked.
- Your permissions.allow rules
- Can't skip an approval while a reviewer can still answer it. After the window, the decision comes back to Claude Code, where an allow rule applies as before.
- If the decision server refuses to answer
- The hook gives no decision, and Claude Code's own permission flow decides. This is a defensive path for an unexpected refusal, not a plan limit.
- What doesn't carry over
- Your allowlist (AgentTrail has no per-pattern exemption, so change the policy on purpose), any action you made weaker than your organization's, and AgentTrail Guard's own decision log. Your AgentTrail Guard files stay on disk, unread in projects AgentTrail OS decides.
- Cursor
- AgentTrail Guard can't stand down in Cursor yet, so setup stops while its Cursor hooks are installed. Remove them first with npx -y @agenttrail/guard@latest uninstall --agent cursor.
The path
Start with AgentTrail Guard. Sign up free for AgentTrail OS when you want a record.
- 1
Protect yourself
AgentTrail Guard. Free, open source, on your machine.
- 2
Remember
A free AgentTrail OS account, still just you: every session searchable, repeat failures captured, your own rules managed from a screen.
- 3
Protect the team
Team: enforcement synced to everyone, each rule backtested on your own history first.
- 4
Oversee
Business: the approval queue with response-time tracking, and an audit trail with export.
“AgentTrail Guard already blocks it. Why sign up?”
AgentTrail Guard blocks the bad action on one machine, with rules you maintain by hand. AgentTrail OS, free for one developer, keeps every session searchable and turns the mistakes your agents repeat into rules. On Team, those rules sync to everyone, proven on your history before you enable them. Approvals and an audit trail sit on top.
AgentTrail Guard stops the mistake. AgentTrail OS remembers it and turns it into a rule.