Introduction

What is Tollgate?

AI agents can read data safely. The problem is write actions — issuing refunds, updating accounts, deleting records. Without guardrails, a hallucinating agent can cause real damage before anyone notices.

Tollgate sits between your agent and your tools. Every time your agent wants to take an action, Tollgate evaluates it against a YAML policy you define. The decision is instant:

• allow — action proceeds immediately
• deny — action is blocked, agent gets a clear reason
• require_approval — action is held, a Slack message goes to your team, and the agent waits for a human ✓ or ✗

How it works

Three steps to go from zero to protected:

1. Register an agent in the dashboard and copy its API key
2. Write a policy — a YAML file that describes what's allowed, what's denied, and what needs a human
3. Wrap your tools with the SDK — one decorator in Python, one function call in TypeScript

Core concepts

Agents

An agent is a registered identity in Tollgate. Each agent has:

• A unique API key (used by the SDK to authenticate)
• An active policy (the rules that govern its actions)
• A full audit trail

Policies

Policies are YAML documents stored in Tollgate. A policy contains a list of rules — each rule matches an action name, optionally filters on payload values, and returns a decision.

version: 1
rules:
  - action: issue_refund
    when:
      amount: { lte: 100 }
    decide: allow
  - action: issue_refund
    when:
      amount: { gt: 100 }
    decide: require_approval
    approvers: ["#approvals"]
  - action: delete_account
    decide: deny
    reason: "Not permitted via agent"
default: allow

Decisions

Audit log

Every action — allowed, denied, or approved — is logged with the agent identity, action name, full payload, decision, timestamp, and who approved it (if applicable).

Next steps