helix --version and mutagent --version to see which builds you have. Inside Helix, /changelog
prints the notes for the recent Helix releases.
Helix 0.3.0 – 0.3.2: MCP, /bug, sub-agent resume, a two-step /build
Helix 0.3.2
/buildruns in two agent steps. First the engineer writes a plan and builds it, tests first; for an agent deployed by API it edits the agent’s definition instead. Then a read-only reviewer checks the result against the spec and the plan. See Build.- A sub-agent stopped for inactivity reports that it failed. It no longer returns the result of its previous run as if it were new.
- The skills and agents that ship with Helix work from any Helix install.
- Resume a sub-agent you started earlier. Its result ends with its id; resuming and messaging it accept that id, the short ref from the agent list, or a unique prefix of at least four characters. See Sub-agents.
- When a sub-agent asks for a model tier you have not configured, you and the model both see a warning that names the model it runs on instead.
- A question from a session started in the background goes to the session that started it. A person there answers it; without one, that session’s agent answers. If nobody can, the background session decides for itself and continues instead of waiting.
- Crash messages and the
/bugnotice name the Helix binary you are running. - Prime mode: a cell that awaits and ends in an
if,tryor block returns that statement’s value.
- Helix now runs on pi 1.0.
- MCP is on. The MCP servers you already configured work as they are, and their tools reach the model directly. Helix adds no MCP server of its own. See MCP servers.
/bugsends your feedback to the Mutagent team, with an optional note:/bug the footer is cut off. Inhelix agentandhelix --prime, which have no feedback, it says so in one line and sends nothing./thinkingoffers every level your model supports, including max. The footer shows the level you picked.- Prime mode runs its code cells in a sandbox. File reads, writes and shell commands from a cell go through the same permission checks and plan mode as any other tool call. Shell output comes back with its exit code, capped at 1 MiB, with the full output saved to a file. Cells can also find and call your MCP tools.
- Agent mode and sub-agents run as agents in their own right, without Helix’s orchestration prompt on top.
- Helix keeps the prompt cache warm across turns and in sub-agents, so repeated context costs less.
- Fixes: turning
/cloudoff brings back the model you were using.helix --prime -pexits once it has printed its answer. Sub-agent transcripts survive compaction and resume.
Helix 0.2.0 – 0.2.2: installer, questions, helix doctor
Helix 0.2.2
- The installer no longer edits your shell profile. It links
helixinto~/.local/bin, creating it if needed, or into~/binwhen only that one is on yourPATH. If the folder is not on yourPATH, it prints the line to add yourself and the full path,~/.mutagent/bin/helix, which always works. See Install Helix. - Helix runs the mutagent CLI only from your installation. When the CLI isn’t installed, Helix tells you how to install it and stops.
- A question Helix asks you ends with a Chat about this row. Choosing it closes the question and returns you to the chat, so you can talk it over first. Answers you already gave are kept.
- Number keys pick options in a question: in a single-choice question a number selects that option, in a multi-choice question it ticks or unticks it.
helix doctornames the providers you are signed in to, or says the saved sign-in is empty or unreadable and how to fix it: type/loginin Helix, or set a provider key such asZAI_API_KEY. It never prints a key.
- Helix moves to the 0.2.0 version line, together with the other Mutagent packages. It is the same Helix as 0.0.38-m0.
Helix 0.0.37-m0 and 0.0.38-m0: /cloud on your repository, stalled responses
Helix 0.0.38-m0
- A model response that stalls no longer hangs Helix. When the provider sends no response headers, or
no content for 120 seconds, the request fails and Helix retries it like any other provider error.
HELIX_STREAM_HEADERS_TIMEOUT_MSandHELIX_STREAM_IDLE_TIMEOUT_MSset the two limits in milliseconds;0turns one off.
- In a git checkout whose origin is on GitHub,
/cloudstarts the sandbox with your repository checked out, on your current branch when GitHub has it. When the sandbox can’t start,/cloudshows what went wrong and the commands that fix it. - In
/cloud, the sandbox’s tool calls show as the same tool cards as local ones, live and after/resume. - Asked to print its own installed files, agent definitions or system prompt, Helix declines in one sentence and offers what it can do instead. Your own project files are unaffected.
- In a question that takes several answers, typing your own answer under Other and pressing Enter checks it and keeps your other choices.
- Evaluation judging is always done by the evaluator agent, so no Gemini key is needed.
Web app Configuration, new LLM providers, GitHub and Slack, Helix 0.0.36-m0
Web app
- A new Configuration page holds the workspace’s LLM Providers, Environments and Integrations, plus the shared Sandbox Providers list. LLM providers moved here from Settings.
- Signing up asks only for your organization’s name; Mutagent creates a workspace named Default workspace in it. The next step, Connect GitHub, can be skipped.
- OpenRouter (
openrouter) and Z.AI (Coding Plan) (zai-coding-plan) can be added to a workspace. The default Z.AI model isglm-5.3.
- Cloud sessions, managed agents, and GitHub and Slack runs are in early access: not every account is let in yet.
- A cloud session stopped after 15 minutes of inactivity wakes up when you send it a message. The web app shows Waking up…, then delivers your message, and the conversation continues in the same thread.
- Connect GitHub and Slack from Configuration › Integrations. A Slack thread keeps one session for up to 1 hour; mention the app in a new message to start another.
- Langfuse and OpenObserve can be connected to a workspace with
mutagent integrations sources add.
- Helix shows its first screen sooner, and internal startup notices no longer appear in the chat box.
/cloudruns the sandbox on a real model: the one given with--model, else the workspace default, else the first workspace model. Turning/cloudon or off no longer changes your saved default model, and/cloud modelworks before the first sandbox starts./cloud stopends the sandbox session again instead of showing an error.- On the Z.ai coding plan,
--provider zai --model glm-5picksglm-5.3-flash. When your plan does not include the model you chose, the turn fails once with one line that names the model and suggestszai/glm-5.3or a plan upgrade. /buildoffers the Mutagent platform as a target.
Helix 0.0.35-m0: faster first start, safe updates
Helix 0.0.35-m0
- Helix starts faster where it has not run before, such as a new cloud sandbox. It loads the extensions and harness it ships with instead of recompiling or downloading them at startup.
- Reinstalling or updating Helix while a session is open no longer deletes the files that session
runs from.
helix updatedeletes nothing; old builds are removed at a later launch once no running Helix uses them. - A cross-session lock problem shows one warning while it lasts instead of one line per tick.
- The installer and
helix updatesend a User-Agent naming Helix, its version, platform and channel.
Helix 0.0.34-m0, CLI workspaces
Helix 0.0.34-m0
/cloudchecks whether cloud sandboxes are enabled for your account before it starts anything. If they are not, it says so in one line and leaves your local session as it was./cloud statusshows whether cloud sandboxes are enabled for your account, or tells you to runmutagent loginwhen your sign-in has expired.
/cloudruns your turns in a cloud sandbox while you keep the local session./cloud stopremoves the sandbox and reportsremoving, thenremoved.- A session is driven by a goal with done-when criteria. The goal cannot be completed until every criterion has evidence.
helix -p "/diagnose …", and any other Helix command run headless, runs a real turn and prints its reply.- After the curl install,
helixworks in the same terminal when~/.local/binor~/binis on yourPATH. - Web search settings with Exa and Firecrawl; Anthropic and xAI models on Google Vertex AI from
/login; plan mode ends with Approve, Keep or Discard.
mutagent loginstores an organization key that works in every workspace you belong to in that organization, for 30 days.mutagent workspaces use <name>selects the workspace every later command works in.--workspace <name|id>overrides it for one command, thenMUTAGENT_WORKSPACE_ID.--org <id|slug|name>checks the key’s organization; a mismatch exits 1 and sends nothing.mutagent login --org <slug>signs in to another organization.- See Accounts, organizations and workspaces.
Session modes
- Cloud sessions have two modes, named the same in the CLI, the SDK and the API: headless (one task, the reply printed) and interactive (a session you keep sending to).
Helix 0.0.32-m0, install, managed agents
Helix 0.0.32-m0
- A sub-agent can be messaged for as long as it exists, including after it finishes; a message to a finished sub-agent resumes it.
- Sub-agents can be named when spawned, and
SendMessage,StopAgentandListAgentsalso reach other Helix sessions on the same machine.
mutagent install helixinstalls thehelixbinary, the same one the curl installer puts in~/.mutagent/bin. The earlier plug-in install into Claude Code and Codex is retired; see Install Helix.
mutagent agent deploy <agent.md>checks, uploads and activates anagent.mdfolder, andmutagent helix agent @<slug> -p "<task>"runs it in a cloud sandbox.agent checkandagent packwork locally, without a cloud sandbox.
Environments, provider mirror
- Environments:
mutagent env setstores the variables and secrets an agent’s tools need, per workspace. Secrets are write-only: they are never printed back. mutagent providers mirrorcopies the LLM provider keys you signed in with in Helix into the current workspace, when you ask it to.