AI agent governance

AI agents at work, with rules that apply before every action.

Claude Code and Codex already run on your team’s machines, with access to code, servers and data. Fleetly puts a company policy in front of every command, edit and MCP call, on the machine itself, and keeps the proof ready for compliance.

Talk to usRead the docs
Applies on the machine

The policy runs where the agent runs, before it acts: in auto mode too, in routines and with the dashboard closed.

Can’t be switched off

IT installs the protection through MDM, and whoever uses the machine can’t remove it. If it drops, the building is told.

Leaves proof

A chained audit log, with a daily copy outside the database and a signed attestation. And all of it streams to your SIEM.

Rules before every action

Blocked commands

What never runs, like git push --force or DROP TABLE: the machine refuses on the spot, in Claude Code and Codex. On Team too.

Two-person approval

Sensitive commands, like terraform apply, and pushes to protected branches only run with two different people approving.

Folders and network

Blocked folders or only allowed ones, and network access only to approved domains.

Hours, models and MCP

Outside working hours, the agent only reads. Only allowed models, only listed MCP servers and a minimum Claude Code and Codex version.

Maximum session mode

Forbid skipping permissions, or the modes that accept on their own as well.

Simulate before it applies

See how many times the policy would have blocked or asked for approval last week. Once saved, it reaches the machines in seconds.

Per-machine policy

The production server can have its own rules, which only tighten the building’s: the stricter one always wins.

Protection people can’t switch off

Managed hooks

One command, or a file pushed through MDM (Jamf, Intune and others), puts the protection in Claude Code’s managed settings.

No shortcuts

You can allow only managed hooks and forbid “skip permissions” on the machine.

Inventory

Every machine, with its Claude Code, Codex and collector versions and its protection: managed, the person’s, or none.

Alert when it drops

Owners and auditors are notified, and the policy can make the machine read-only until the protection is back.

Code and secrets where they belong

Secrets masked at the source

Keys, tokens and passwords become ••• on the machine itself, before any conversation or output reaches the hub.

Code stays on the machine

Agents run where they are, with your accounts. Claude Code, Codex, Antigravity and GitHub Copilot credentials are never read.

The machine decides what it shows

Only work folders show up for the building. The hub can’t widen that.

Per-machine privacy

Whoever looks after a machine chooses whether teammates act as their role allows, only read, or don’t see the sessions at all.

Who gets access

Roles

Owner, operator, viewer and auditor. The auditor reads the audit log and reports without seeing conversations, and doesn’t take a paid seat.

SSO

Entra ID, Okta, Google Workspace, Keycloak or any OIDC provider, with a verified domain. SSO can be required.

SCIM

Your company directory adds and removes people. Removing someone ends their sessions right away.

Access reviews

Every period, the owner confirms who stays; whoever is marked leaves the building.

Audit log and proof

Everything audited

Sign-ins, messages, approvals, commits, pushes and settings changes, with the person, address and time.

Agent trail

Every agent action on the machines, with who authorized it and whether it came from the dashboard or the machine.

Who saw what

What each person opened from the building’s machines: conversations, diffs, attachments, folders and searches.

Integrity attestation

Records are chained, and every day a signed copy goes to locked storage outside the database. Any change, even by someone with database access, stops matching.

In your SIEM and your reports

Splunk, Datadog and webhooks

The audit log and agent trail stream every 30 seconds, without gaps. Webhooks are signed.

Read API

Pull the audit log and trail with a read-only service token, scoped to its building.

Key events

Policy changed, protection dropped, budget reached, break-glass access: just what matters, good for Slack or Teams.

Cost and budget

Estimated and actually billed cost, per project and machine, with a monthly budget and an optional lock when it’s reached.

What each plan includes

Team

Roles, audit log, command blocking, masked secrets in conversations and per-machine privacy.

Enterprise

Everything on this page: full policy, managed protection, SSO and SCIM, auditor, reports, agent trail, attestation and integrations.

Govern the agents, don’t watch the people

The numbers are per project and per machine, never a ranking of people. The policy says what agents may do; nobody gets a screen to measure a coworker.

Documents for your review

Data processing agreement (DPA), subprocessor list, backup and recovery, incident response and a completed security questionnaire. Ask at admin@fleetly.tools.

Governance questions

What is AI agent governance?

Deciding what coding agents, like Claude Code and Codex, may do on company machines, making sure those rules always apply and keeping proof of what happened. In Fleetly, that’s the building policy, managed protection, roles, SSO and the audit log.

Do the rules apply when nobody is watching the dashboard?

Yes. The policy runs on the machine itself, before every command, edit or MCP call, in auto mode too, in routines and with the dashboard closed.

Does it work with Codex?

Yes. Blocks, approvals and the other rules apply in Claude Code and Codex. In Codex, Fleetly’s hooks need to be approved once in Review hooks.

Does Fleetly see our code?

Code stays on the machines. The hub carries session state and whatever someone opens in the dashboard, always encrypted, with secrets masked before they leave the machine.

How do I show this to audit or compliance?

With the signed integrity attestation, reports in PDF and CSV, continuous streaming to your SIEM and the document pack (DPA, subprocessors, backup, incidents and security questionnaire).

Is this for monitoring employees?

No. The numbers are per project and per machine, never a ranking of people. The policy limits the agents, not the people using them.

Want to see the policy running on your machines?