Documentation
How Helix works
The loop end to end, then the parts you configure. Each section says what actually happens.
The loop, in six steps
Every run follows the same shape. Nothing here is a mode you turn on.
You state an outcome
Plain words. That is the whole input — no schema, no prompt syntax, no tool selection.
Add a rate limit to the public API and prove it works.Helix drafts a plan
The goal is broken into steps, and each step is routed to the specialist whose tools fit it. You can read the plan before the work starts.
A sandbox is provisioned
The run gets its own machine with its own filesystem and network namespace. Commands and files happen there, not on your machine.
The crew works
Specialists execute steps and hand work between each other. Every handoff is written into the transcript in that specialist's own voice.
It verifies
A change with no check is reported as unverified, not as done. The summary is built from tool outcomes, not from the model's account of them.
You keep the record
The transcript, the artifacts and the workspace all persist. The machine is parked between messages, so tomorrow's follow-up finds its files.
The crew
Five specialists ship with every workspace. Each has its own tools, so work is routed rather than piled onto one generalist.
Marcus Chen
Senior Software Engineer. Edits, tests, builds, refactors.
Priya Nair
Cloud Engineer. Services, deploys, logs, capacity.
Diego Ramos
Data Analyst. Queries, CSV and spreadsheet work, charts.
Sofia Lindqvist
Research Analyst. Web search, reading, citations.
Amara Okafor
Virtual Assistant. Writing, summaries, scheduling.
Your own
Describe a specialist in a sentence and it joins the roster.
Every run gets its own machine
Isolation is the reason a run can be trusted with a real shell.
A real machine, not a simulation
The sandbox is a real Linux environment. A run installs a dependency, builds the project and runs the tests exactly as you would — because it is doing exactly that.
It cannot read another workspace's files. Loopback inside it is the sandbox, not your host, which is what makes a server started there safe to expose through a signed preview URL.
$ pwd /workspace $ node --version v22.14.0 $ npm test 42 passing (1.2s) $ echo $DISPLAY :1 # a real X server, so a GUI tool works tooIt checks before it says done
The claim being made is not “the model said so”. It is that a check ran.
verified
A check ran after the change and passed. The evidence names which check.
unverified
The change landed and nothing proved it. Reported as exactly that, with the gap written down.
awaiting_acceptance
Finished, and a person should decide. The reason completion was withheld is on the row.
What a run can call
Seventy tools, grouped. A run is given the ones its goal needs rather than all of them, so the manifest stays small and the model is not invited to reach for something irrelevant.
Files
read_file, write_file, edit_file, apply_patch — every write read back to confirm it landed.
Shell
run_command, start_service, sandbox_jobs — with the process's own log tail on a crash.
Code intelligence
find_symbol, find_references, call_graph, repo_map.
Git and GitHub
git_show, open_pull_request, github_actions, check_merge_conflicts.
Browser and desktop
browser_use, computer_screenshot, computer_click, computer_type.
Data
create_document, export_artifact, read_document — PDF, XLSX and DOCX with no extra install.
Research
search_codebase, search_platform_docs, npm_registry_lookup.
Delivery
publish_artifact, get_preview_url, request_env_values.
The controls you keep
Autonomy is a level, not a switch. It says how much your crew may do without asking.
approval_required
Nothing runs without asking. Every action is proposed first.
supervised
Reads and reversible work run. Writes and external effects ask first.
autonomous
Everything runs except actions above your risk tolerance.
Questions people ask
Do I need to know how to code?
No. The input is a sentence describing an outcome. The plan, the tools and the checks are the platform's job.
What happens if a run gets something wrong?
You see it in the transcript while it happens, and you can stop it. Every finished task carries its evidence, so you can tell what was actually verified from what was merely claimed.
Can it touch my production systems?
Only through credentials you connect, and only within the autonomy level you set. A destructive action asks regardless of the level.
Is my code used for training?
No. Your workspace, your transcripts and your artifacts are yours; they are not training data.
What happens when a usage window runs out?
The run pauses and the message names the reset time. Nothing is lost — send another message once the window rolls.
See it work on your own problem
Describe an outcome and watch the plan, the work and the proof arrive in one transcript.

