Dev Tools & IDEs · Submission selection

Kaeso

Run an open-source AI gateway with provider routing, project keys, request logs and configurable budgets.

Hands-on checks 15 November 2026. Keep for developers who can operate a self-hosted gateway. No numerical quality rank assigned.

Why this earned a place

The local installation routed a real HTTP request through a controlled OpenAI-compatible mock, returned streaming output and enforced key-level limits before contacting the upstream. That is a useful, concrete infrastructure workflow.

Access and cost boundaries

MIT-licensed software. Hosting, infrastructure operation and model-provider usage remain separate costs.

Limits to consider

Requires deployment, databases and ongoing operation. Our upstream was a synthetic mock, not a paid model provider. Dollar-denominated spend limits, provider failover, ClickHouse request logs, SSO, load capacity and security were not independently validated.

Data handling

Gateway logs can include request content and media when enabled. Configure retention and access deliberately; self-hosting alone is not a security guarantee.

What we checked

Installed release 0.6.0 in an isolated local Docker environment with PostgreSQL and Valkey. Tested chat-to-Responses translation, streaming, usage fields, invalid-key rejection, rate limiting, hard token-budget rejection and restart recovery. A hard token budget also refused a request while its counter service was unavailable, without calling the mock provider. Automated tests used synthetic content and no paid model calls on 15 November 2026.

Sources and evidence

Hands-on test record · Original selection record · How we review

Free editorial selection. No payment, service purchase or badge was required. This page does not claim a comparative winner, security certification or exhaustive product test.

Fewertools editorial selection · Suggest a correction