TollgateThe AI gateway for coding agents

One Go binary that controls cost, rate limits and failover for the Anthropic and OpenAI APIs.

Hard budgets

Set a budget for each tenant, each month and each agent session. A hard budget refuses the request with status 402 before the provider gets it.

Rate limits and budgets

Failover before the first byte

Ordered targets, retries, a retry budget and circuit breakers. Tollgate moves to the next target before the client sees an error.

Reliability

Built for agent sessions

Session budgets, sticky routing that helps preserve the warm prompt cache, an exact response cache and personal data redaction.

Agent features

Metrics, traces and cost

Prometheus metrics, OpenTelemetry traces with one span for each attempt, and one JSONL usage event with the cost of each request.

Observability

Dashboard in the binary

Overview, sessions, providers and keys, in light and dark mode. The admin port serves it. No separate process is necessary.

Admin API and dashboard

Native API formats

Anthropic /v1/messages and OpenAI /v1/chat/completions keep their format. Streams pass through byte for byte, unless redaction or a budget cutoff changes them.

Configuration

See it work

The 81-second tour explains request controls, failover, circuit breakers, budgets, sessions, traces, and the dashboard. Animated examples use Tollgate behavior and local fake-provider output. Screenshots use synthetic data. See how the video is made.

Start in one command

For a local dashboard with synthetic traffic and no API key, clone the repository and run:

make dashboard-demo

Open http://127.0.0.1:9090. You need Go, GNU Make, and Node.js. See the quick start for a first API request.

Install

go install github.com/ykakade/tollgate/cmd/tollgate@latest

Release binaries for Linux are on the releases page. Continue with the quick start.

The dashboard

Tollgate dashboard with spend, request totals, cache use, and route costs

The screenshot uses synthetic demo data. Run make dashboard-demo to show it on your computer.