Navige: Cut AI agent costs and control runaway spending

Navige works as a proxy in front of model calls, so it can change a step before it runs rather than report the bill afterward. Setup is described as one line of config, and it covers Claude, GPT, Gemini and Bedrock traffic. Its model is a ceiling: calls are only moved down to cheaper models, never up. If a smaller model fumbles a tool call, the request re-runs on the original model and the agent sees only that result. Pricing is the main open question in the launch thread. Watchtower, the monitoring tier, is free. On paid plans, customers keep four parts of every five of the measured savings Navige can show, and Navige keeps one, with a $9 monthly floor. Claude Code is supported through the Anthropic proxy, though some optimisations are deliberately restricted there to avoid interfering with how it works. For founders, agent spend tends to grow through long contexts, retries and loops, and few teams can tie it to specific jobs. A gain-share fee ties vendor revenue to results, which lowers adoption risk. The catch is that savings figures depend on Navige's own measurement, so teams should check how measured and estimated numbers are separated before trusting them.