What Is Gatekeeper

Gatekeeper lets your AI agents work with your real systems, like HubSpot, GitHub and Slack, without handing them your credentials or unlimited access. Every action an agent tries to take is checked against rules you set before it runs, and every request is recorded.

Why it matters

Connecting an AI agent straight to a service usually gives it the same access as the person who set it up. That's often everything. The API key or token behind it tends to sit in a config file the agent can read.

So one misread instruction, or a malicious prompt hidden in an email or document, can become a real action against your systems. Gatekeeper puts a checkpoint in between.

How it works

Gatekeeper sits between your AI client — Claude, ChatGPT, Cursor, or any MCP client — and the services it uses. Instead of connecting your client to each service directly, you connect it once to your Gatekeeper gateway. Every request then follows the same path:

1. Your agent asks to take an action, such as updating a HubSpot contact.
2. Gatekeeper checks it against your policy: which service, which tool, and what it's trying to send.
3. Gatekeeper decides: allow it, deny it, or hold it for someone on your team to approve.
4. If allowed, Gatekeeper makes the call using your stored credential and passes the result back to the agent.
5. The request is recorded in your audit log, whatever the outcome.

Your agent sees the same tools it always did. It just can't go around your rules.

How Gatekeeper protects you

Your credentials never reach the AI

When you connect a service, its credential — an API key or OAuth token — goes straight to Gatekeeper and is encrypted immediately. It's only unlocked inside a secure, isolated environment when an approved request needs it.

It's never shown to your agent, never written to logs, and never stored in your AI client's configuration. If an agent is compromised, there's no key for it to leak. See How Gatekeeper Protects Your Credentials.

Control over individual actions

Access isn't all or nothing. Gatekeeper evaluates each tool call on its own, so you can let an agent read contacts but not delete them, or comment on a pull request but not merge it. You can also switch off specific tools for a service entirely.

Rules, not guesswork

Your policies are fixed rules you write, not an AI model judging whether a request looks safe. The same request gets the same answer every time, and there's no prompt that can talk its way past them.

You choose what happens when no rule matches. During setup you pick a default posture — Allow, which lets unmatched requests through while you observe what your agents actually do, or Deny, which blocks anything not explicitly permitted. Allow is the recommended starting point, but it does mean requests you never explicitly permitted will go through until you add rules. See How Policies Are Evaluated.

A person in the loop for sensitive actions

For high-stakes actions like deletions, bulk updates, or anything touching money, you can require approval. Gatekeeper holds the request until someone on your team approves it, and nothing runs until they do. See Approving a Flagged Action.

A record of every request

Every request through Gatekeeper is logged: when it happened, the service and tool, the details sent, and the outcome, including requests that were denied. Use it to confirm your policies work, investigate an issue, or answer a compliance question. See Reviewing Your Audit Log.

Important: keep every request going through Gatekeeper

Gatekeeper can only protect requests that pass through it. If your AI client keeps a direct connection to a service alongside its Gatekeeper connection, your agent can take the direct route. That skips your policies entirely and leaves no audit record.

When you connect a service through Gatekeeper, remove any direct connection to that same service from your AI client.