Introduction
Gasket runs AI agents for your application and controls what they do.
Gasket is infrastructure for putting AI agents inside your product. You configure an agent in the dashboard (instructions, tools, and the policy that governs them), create an API key, and your server calls the agent over HTTP. Gasket runs the model, checks every tool call against your policy, pauses for a person when a rule says so, and records every request, decision and result.
Two ways to use it
- Run agents on Gasket:
POST /v1/agents/:id/runswith a message. Gasket calls the model, runs your tools through the policy, and returns the reply. This is the fastest way to add a chatbot or assistant to an app. - Control agents you already run: before a consequential step, your own agent code calls
POST /v1/actions.Gasket answers authorized, denied or pending_approval. Or connect an AI client like Claude Code over MCP.
What's included
| Piece | What it does |
|---|---|
| Agents | Instructions, model (Claude Haiku 5.5), tools, policy, environment and status. |
| Runs | One call to an agent: model turns, tool calls and the reply, metered by tokens. |
| Tools | Your HTTPS endpoints. Gasket calls them only when the policy allows, with a signed request. |
| Policies | Rules that allow, deny or require approval per service.operation, with amount thresholds. |
| Approvals | Held actions wait for a person in the dashboard; runs resume when approved. |
| Requests, activity, audit | Every API request with its response, every event, every decision. |
| Webhooks | Signed events pushed to your endpoints, retried with backoff. |
| Billing | Prepaid credit, $0.50 free, usage ledger, monthly spend limit. |
NoteStart with the Quickstart: a working agent and your first request in about five minutes.
Something wrong or missing? Tell us · Stuck? Help center