Skip to main content
GET STARTED

Welcome to Rungate

Rungate is a governance control plane for AI agent runs — budgets, policies, and approval gates applied to the complete workflow from first model call to final tool invocation.

What Rungate is

Rungate is an HTTP proxy that sits between your AI agents and LLM provider APIs (OpenAI, Anthropic). Your agent points its existing provider SDK at Rungate instead of the provider directly. Rungate forwards each request — preserving the original request and response format — and enforces governance along the way.

The unit of governance is a run: the complete piece of agent work, from first model call through every tool invocation, retry, and approval gate, to the final output. Budgets, policies, and audit trails apply to the run as a whole — not the individual API call.

The shortest version. Sign up at rungate.dev/signup, point your OpenAI / Anthropic SDK at https://api.rungate.dev/v1, use your rg_agt_* token as the API key, and you have a governed proxy in front of every call.

What it gives you

  • Run-level budget enforcement. Set a hard ceiling per run (or per agent, or per hour). Cumulative spend is tracked across every call. When the cap is hit, the next call returns 402 Payment Required with a structured context object so your agent knows exactly why and what to do next.
  • Policy controls. Allowed model lists, blocked tools, rate limits, destructive-action gates. Declared once, enforced server-side, applied to every call in the run.
  • Human approval gates. When a tool call matches a rule that requires human review, Rungate returns HTTP 202, pauses the run, and waits. The agent retries the same request later; on approval the retry succeeds.
  • Cross-provider failover and translation. Configure a fallback provider and Rungate transparently re-emits a failed request — across vendors, with format translation. Agents see the response in the format they sent.
  • Full audit trail. Every model call, tool call, retry, approval, and governance event is reconstructable per run.

Choose your path

Rungate is open source under Apache 2.0 and runs in two ways:

  • Managed cloud — sign up at rungate.dev/signup, get tokens immediately, use the free tier (500 successful steps / month) to evaluate. No credit card.
  • Self-host — Apache 2.0, one process, one SQLite file, unlimited usage, no telemetry. The repository is currently private during early access; request source access by emailing [email protected]. Setup details on the Self-host page.

Where to go next

  • Quickstart — sign up, get tokens, make your first call, in five minutes.
  • Runs — the central concept. How to group calls, when runs auto-close, how long-horizon runs work.
  • Policies & budgets — what you can enforce.
  • Approval gates — the HTTP 202 pause-and-resume flow.
  • Integrate your agent — code snippets for OpenAI, Anthropic, and curl.

Reference

  • Agent reference — single-page, dense, intended for AI agents to read end-to-end. Useful for humans too if you want everything on one page.
  • OpenAPI 3.1 spec — every endpoint, schema, and error code.