Skip to main content
Helix, Mutagent’s coding agent that evaluates and fixes AI agents (What is Helix), connects to the tools around your agent in four ways: it reads traces from wherever your agent runs, it writes approved fixes back to your agent’s source, it takes requests through channels like GitHub and Slack, and it works with agents that run in the cloud.
Helix on your machine reads traces where they are; it does not store a copy. A trace source you connect to your Mutagent workspace (mutagent integrations sources add) is different: there, Mutagent pulls the traces into your workspace.

Trace sources

Where your agent’s runs come from: local transcripts, JSONL files, Langfuse, or OpenTelemetry (OTLP/JSON) span files. Dash0: Coming soon.

Apply targets

Where approved fixes land: markdown agents or agent code, as a GitHub pull request.

GitHub and Slack

Pull requests for fixes. Start runs from issues, mentions and CI, or from a Slack mention.

Cloud deployment

Mutagent managed agents, and agents on Paperclip, Mastra Cloud, Vercel or the Claude Agent SDK.

​
Local and hosted

  • Local integrations need no Mutagent account: Claude Code, Codex and Helix transcripts, JSONL files, and pull requests from a checkout you own, with your own git credentials.
  • Hosted integrations use credentials for another service: an observability backend for traces, whose keys Helix reads from environment variables; or the Mutagent GitHub App and the Mutagent Slack app, connected once for your organization in the web app (Configuration › Integrations) or with mutagent gateway connect.

​
Example: evaluate from Langfuse, fix by pull request

1

Install Helix with a model key

Follow Install Helix. helix --list-models must list at least one model.
2

Write the config

Save this as .mutagent/config.yaml in your project:
.mutagent/config.yaml
3

Export the Langfuse keys

Helix reads the Langfuse keys from these fixed variable names (or from the project’s .env):
4

Check the setup

/status shows the current stage and anything still missing. It needs a working model key.
With this config, /evaluate scores the agent on its production traces in Langfuse, and an approved fix from /optimize arrives as a pull request that edits .claude/agents/. Helix opens the pull request with the git and gh credentials you already push with.

Supported integrations

Every LLM provider, sandbox provider, trace source, apply target and channel, with its status.