Skip to main content
Check here whether Mutagent works with your stack. Each row links to its setup page.

​
LLM providers

You bring your own keys. An LLM provider added to a workspace belongs to that workspace only; two workspaces never share one. Add them in the web app under Configuration › LLM Providers or with mutagent providers add, or copy the keys you signed in with in Helix using mutagent providers mirror. Every entry is tested against the provider before it is stored, and a stored entry can be re-tested with mutagent providers test <id>. mutagent providers list --models shows the models each entry offers. OpenAI, Anthropic, Google Gemini, Moonshot AI, Z.AI (GLM), DeepSeek and xAI also take an optional base URL override and a model allow-list. For Z.AI (GLM), the default model is glm-5.3. Z.AI (Coding Plan) offers the GLM models your plan includes. LLM provider setup lists every field.
Helix on your machine does not need any of these: it calls the provider directly with the key in your environment or the one you set with /login. Workspace LLM providers are used by the hosted parts, such as cloud sessions and managed agents.

​
Sandbox providers

A sandbox provider runs cloud sessions and managed agents: an isolated machine with Helix and your agent’s tools on it. Cloud sandboxes, and the GitHub and Slack runs that use them, are not open to every account yet. They open to every account after production acceptance. LLM providers and Environments are per workspace because they hold your keys and your secrets. The sandbox provider holds neither, so one list serves every workspace; the keys and Environment a session needs are passed in from the workspace that launches it. mutagent sandbox providers lists the sandbox providers, and mutagent sandbox presets lists the presets: named sandbox definitions a cloud session can be launched from. See Sandbox providers.

​
Trace sources

Helix reads traces where they already are. It does not copy them to Mutagent. Local sources are described in Local trace sources. Langfuse and OpenObserve can also be connected to a workspace once, with mutagent integrations sources add, so Mutagent pulls the traces for you instead of Helix reading them from your machine.

​
Apply targets

After you approve a fix, Helix writes it to a target. Every target lands the change as a GitHub pull request on your repository; nothing is written to your default branch. With no target configured, every stage is report-only.

​
Channels

Connect them in the web app under Configuration › Integrations (each has a Connect button), or with mutagent gateway connect github and mutagent gateway connect slack. Each app is installed once for your organization. The repositories Mutagent can use are the ones you give the Mutagent GitHub App access to on GitHub. GitHub and Slack connect through the gateway, which starts each run in a cloud sandbox.