We've found that power users generally want to hand a CLI/MCP to their agents and get started from the terminal. They rarely want to log in and use the web UI themselves. We think developer tools must be agent-friendly to remain relevant.
Here are some common findings that Gauge identifies:
- agents get confused with CLI/SDK behavior and start source diving - interactive commands block agents - agents use deprecated/old CLI commands - agents don't use Skills when you'd expect - agents confidently assert limitations that don't exist - inconsistent terminology confuses agents - agents can't figure out how commands fit together into a workflow
Locally, the Gauge CLI uses Docker containers to run agent harnesses (you can bring your own subscriptions). It finds where agents get confused or stuck and tests changes using your pre-release artifacts (Skills/CLI etc). We normalize events from Claude Code, Codex, Pi and OpenCode into standardized transcripts, and store them locally, alongside structured information in SQLite so that your agent can easily analyze the data. This enables automatic hillclimbing toward a great Agent Experience.
We also offer hosted compute that enables more features, built on our custom Gateway (for credential brokering and transparent, per-session response rewriting). Hosted compute enables many more concurrent sessions, easier collaboration with teammates, a GitHub App for PR reviews, sessions with logged-in MCPs/APIs, and most importantly it lets you close your laptop while the sessions actually run.
You don't need a Gauge account to run the full loop end-to-end though. We expect that local execution with the CLI is enough to find and fix several agent experience issues in a typical project.
You'll be able to validate fixes for these issues locally, before publishing new versions (we inject pre-release Skills and code artifacts into the sessions).
We would love your feedback!