Skip to main content
The Langy home with the chat panel open and suggested prompts

The Langy panel open on the LangWatch home. Ask it a question and it starts reading your traces.

Langy is a full coding-agent harness running in a sandbox, with access to everything you grant it in your LangWatch account, so it fully understands your agents, evals, and prompts. Being a coding agent, it connects to your GitHub, runs code, experiments with improvements, and contributes them back as pull requests. With a CI pipeline that reaches your other systems, Langy runs its evals against your real agent pipeline, however complex it is. Langy uses your own model accounts and API keys. Every turn it takes is itself a traced, costed run in your LangWatch project, so you watch Langy work with the same tools you use to watch your own agents.

​
Where to go next

Getting started

Enable Langy, pick a model or sign in with Codex, and review the first pull request.

How Langy works

The mechanism: a gVisor-sandboxed opencode worker whose only interface to LangWatch is the langwatch CLI.

Ask Langy

The chat panel and the kinds of questions Langy answers, from regressions to cost spikes.

Models and Codex

The model picker, bring your own API key, and signing in with your OpenAI account for Codex.

Security and the sandbox

What the sandbox can and cannot reach, and where the boundaries sit.

Self-hosting Langy

Ships enabled in the Helm chart, sandboxed with gVisor by default.