What is GaiterGuard?
GaiterGuard is an intercepting API gateway that enforces human-in-the-loop (HITL) authorization for autonomous AI agents. Agents never hold production credentials. High-impact actions require explicit human approval through an out-of-band dashboard. The trust boundary is enforced by the gateway — not by the agent.Trust boundary principle: Your AI agents never see real API credentials. GaiterGuard injects credentials at request time and blocks risky operations until a human approves them.
How it works
GaiterGuard sits between your AI agents and external APIs, evaluating each request through an LLM-powered risk assessment:1
Agent sends proxy request
Your agent calls
POST /proxy with an Agent-Key, target URL, HTTP method, and a natural language intent describing what it wants to do.2
LLM risk assessment
GaiterGuard’s risk assessor reads the API documentation, your custom rules, and the agent’s intent to compute a risk score (0.0–1.0).
3
Low risk: forward immediately
If the score is below your threshold (e.g., 0.5), GaiterGuard injects credentials from the encrypted vault and forwards the request to the target API immediately.
4
High risk: require approval
If the score meets or exceeds the threshold, GaiterGuard returns
428 Risk-Blocked with an action_id. The request is queued for human review.5
Human approves or denies
You review the request in the dashboard, see the full context (intent, payload, API docs), and approve or deny with one click.
6
Agent polls and executes
The agent polls
GET /status/{action_id} every 5–10 seconds. Once approved, it calls POST /proxy/execute/{action_id} to execute the request with real credentials.Key features
Secret vault
Credentials encrypted at rest with AES-256-GCM. Never returned in API responses or exposed to agents.
Agent-Key authentication
Scoped, revocable keys. One key per agent. Agents never see service credentials.
LLM risk assessor
Evaluates requests against API docs and custom rules (e.g., “any Stripe charge > $100 requires approval”).
Intent integrity check
LLM compares the agent’s stated intent against the actual payload to detect mismatches or suspicious behavior.
Approval queue
Dashboard shows full request context. Approve or deny with one click. All actions are logged.
Global kill switch
Block an entire agent session instantly by revoking its
Agent-Key or setting isActive: false.Idempotency
Duplicate requests are deduplicated via
Idempotency-Key headers. Prevents accidental double-execution.TTL cleanup
Approved-but-unexecuted requests expire automatically after a configurable window (default: 1 hour).
Tech stack
GaiterGuard is built with modern, performant tools:Use cases
Customer support agents
Customer support agents
AI agents that need to refund charges, cancel subscriptions, or modify user data. Low-risk queries (“get invoice”) pass through instantly. High-risk mutations (“refund $500”) require approval.
Infrastructure automation
Infrastructure automation
Agents that provision cloud resources, modify DNS records, or deploy code. Safe reads (“list instances”) auto-approve. Destructive operations (“delete database”) require human confirmation.
Financial operations
Financial operations
Agents that interact with payment processors, accounting systems, or banking APIs. Small transactions may auto-approve based on your rules; large transfers always require approval.
Multi-tenant SaaS tools
Multi-tenant SaaS tools
Agents that perform actions on behalf of multiple customers. Each service can have isolated credentials, and risky cross-tenant operations are blocked until reviewed.
Next steps
Quickstart
Get GaiterGuard running in Docker and make your first proxied request in under 10 minutes.
Installation
Detailed setup instructions for Docker, local development, and production deployment.
Architecture
Understand how GaiterGuard works under the hood: request flow, risk scoring, and approval logic.
API Reference
Complete API documentation for agent-facing and dashboard endpoints.