Skip to main content
Research Preview. Commands and output can change between releases. Run helix --version to see the build you have. The loop runs end to end today.
Helix is a coding agent that builds and improves AI agents. It runs in your terminal and takes an agent through the five stages of the agentic development lifecycle (ADLC): spec → build → evaluate → diagnose → optimize. Optimize loops back to Evaluate: fix, re-evaluate, repeat. Every change waits for your approval and is re-checked against your agent’s real runs (its traces). Shipping is a separate step after the loop.

​
Point it at an agent

Point Helix at an agent, one you’re building or one already running, and it works through the stages: it writes a spec, builds the agent, scores it against your real traces, finds out why it failed and proposes fixes, then applies the fixes you approve and re-evaluates. Nothing changes until you approve it. See how the loop works for the full mechanism.
No traces yet? Building something new is fine — start at Spec and Build. Evaluate and Diagnose start working once you ship and your agent begins logging runs.

​
Install it

Helix is one program, helix, with everything inside: sub-agents (helper agents it starts for parallel work), the trace viewer, and its other modes. You don’t need Node, npm or a source checkout.
The installer puts the binary at ~/.mutagent/bin/helix and links it as ~/.local/bin/helix, so helix works in the same terminal. It does not change your shell configuration. Helix needs a model provider key. Export one before you run it, for example export ANTHROPIC_API_KEY=sk-ant-... (or run /login inside Helix). Check that the provider works with helix --list-models: it lists the models your key can use. Then run helix in your project folder. Bare helix needs an interactive terminal. mutagent install helix installs the same binary, for when you already use the mutagent CLI. To use Helix from a Claude Code or Codex session, have that agent run helix -p "<prompt>": it sends one prompt, prints the answer and exits, with no interactive screen (headless).

Install Helix

Install options, updates, and removal. Next: the Quickstart, then your first loop.

​
Drive it in plain English

You talk to Helix in plain English. Describe what you want, for example “Evaluate the Refund Processing agent against last quarter’s disputes and show the pass rate per policy rule”, and it picks the stage that does that job. Nothing runs until you ask, and nothing changes until you approve. helix also has two lighter modes, agent mode and Prime, for when you only want a coding agent: See Three ways to run Helix for what each gives you and what to type.
Which model providers, trace sources (where your agent’s runs are read from) and apply targets (where approved fixes are written) does Helix work with? See Supported integrations.