Skip to content
LaunchHelm
Documentation — The assistant

The assistant

The assistant answers questions about your runs from inside LaunchHelm: why a metric moved, what changed before it, how runs compare. It reads what you can read and shows the charts and numbers its answer rests on. It can propose a change for you to send, and, if you allow it, make changes itself.

Ask

Open Ask in the top bar. The panel docks beside the page on a wide screen and fills the screen on a phone.

The panel starts with what you are looking at: the run, its project, or a chart. Each shows as a chip above the question box. Remove a chip to leave it out of the question. Other ways in:

  • The command palette. Type a question and choose Ask: at the end of the list.
  • An alert on a run. What happened? asks about the alert.
  • A zoomed board. Explain this range asks about the span you zoomed to.
  • An investigation. Suggest hypotheses asks for explanations worth checking against the evidence.
  • A chart's ⋯ menu. Choose Ask about this chart. The question then carries the chart, what it reads, and the window on screen.

Press Enter to send, or Shift+Enter for a new line. While the assistant works, its answer appears as it is written, and the send button becomes Stop.

What an answer holds

  • Text in Markdown, with tables and math.
  • Tool calls, folded to one line each, saying what was read. Open one to see what came back.
  • Charts. A chart is a real board panel, drawn from your data now. On a board you can edit, Add to board puts it on the page you are on, as one change you can undo. Save the board to keep it.
  • Tables comparing runs, settings or results.
  • Buttons that do what a menu or the command palette would: open a run on one of its tabs, explore a metric, open the alert composer. They run only when you press them, in every mode, because they act on your screen.

Numbers in an answer come from what the assistant read in that conversation. When the data was bucketed or cut short, the answer says so. A value of a metric, such as the latest loss or its value at a step, is a live value: it is read from your data wherever the answer is shown, as a text panel's live values are, and hovering it says which run and which reading it is.

What it can read

The assistant acts as you. It sees what you can see:

  • documents and their settings;
  • metrics, read with queries;
  • logged samples;
  • events, commands and rule executions.

Text inside your data, such as a run's name or a logged sample, is treated as data, never as instructions to the assistant.

Modes

The mode menu beside the conversation's title sets what the assistant may do in that conversation. You can change it at any time.

  • Ask reads and answers. It can still propose a change.
  • Propose proposes changes. A proposal shows as a card; Review… opens the run's request form filled in, and nothing is sent until you send it. Asked to arrange the board on screen, it proposes moves, sizes and new panels; Show on board puts them on the board as one change, which you save to keep or undo.
  • Full permission also makes changes itself, as you, and shows each one as it does it, with Follow to watch it apply.

Alerts work the same way. Ask to be told when something happens, and the assistant drafts an alert: a metric, when it fires, and how loud. Review… opens the alert composer filled in, where you can check it against history before you save it. In Full permission mode it creates the alert itself, telling you in the app.

If you couldn't make a drafted change yourself, because you don't have permission to, the card says so instead of offering it; someone who has permission can make it.

Changes to investigations work the same way. In any mode, the assistant can draft a new investigation, a hypothesis, a hypothesis's status, or evidence for or against one; the card's Apply makes the change as the investigation page would. If someone changed the investigation since the draft, the card says so instead of applying it; ask again to redraft. In Full permission mode it makes these changes itself.

In every mode, a change that needs approval waits for another member, and the assistant never approves anything. It can't create keys, change who is in the organization, or change billing.

Research

When your deployment turns them on, the assistant can search the web and a connected research server, such as a paper index. Each search shows as a tool call with its query, and the pages an answer cites are listed under it. Only the assistant's search queries leave LaunchHelm, and it is told not to put your runs' names or numbers into them unless you wrote them in your question. A paper's claim is set beside your charts as a hypothesis to check, not as a diagnosis.

Conversations

Conversations are private, and every one is kept: come back to any of them from History, which lists yours and those shared with you, newest first, with where each started. Search their titles, and page back with Show older.

To keep one close at hand, save it: the star beside it in History, or Save in its ⋯ menu. Saved in History lists only those. Saving is yours alone; you can also save a conversation someone shared with you.

The panel keeps the conversation you have open as you move between pages, and opens it again when you reload. ⋯ beside a conversation's title also renames it or copies its link: the link opens the conversation in the panel, on the page it started on, for anyone who can read it.

Delete… in the same menu deletes a conversation you started, with everything in it, once you confirm. It can't be undone: people it was shared with lose it too, and it leaves any investigation it was attached to. The assistant's usage it counted stays counted. Stop an answer before deleting its conversation.

A conversation started in a project can be shared with that project: press Private to switch it to Shared. People who can read the project can then read the conversation, but only you can continue it. Press Shared to make it private again.

Under a conversation of yours, Attach to an investigation… adds it to an open investigation of its project as evidence, sharing it if it is private. Turn into an investigation asks the assistant to draft one from the conversation: the question you asked and the hypotheses discussed. Open it from the card, then Attach this conversation to keep the two together. The investigation lists it under Conversations, and Detach takes it off. Making the conversation private detaches it too. A run's Details tab lists the conversations started in it that you can read; open one to read it in the panel.

The allowance

Each plan includes the assistant, with a monthly allowance of model tokens for the organization. The meter under the question box shows how much of this period's allowance has been used. When it runs out, conversations stay readable and new questions wait until the period resets.

Running your own deployment

The assistant needs a model. Without one, conversations stay readable but no question can be asked. Four settings choose it:

  • LAUNCHHELM_ASSISTANT_PROVIDER: anthropic (the default), openai for the OpenAI Responses API, or openai-chat for the OpenAI Chat Completions API. The OpenAI ones also cover every server compatible with them: hosted providers, gateways and routers, and local model servers such as vLLM, Ollama or LM Studio. Prefer openai where the server offers it; use openai-chat for one that has no /responses.
  • LAUNCHHELM_ASSISTANT_BASE_URL: the server to call, when it is not the provider's own API. With the OpenAI ones, give the API root, such as http://localhost:11434/v1; requests go to its /responses or /chat/completions.
  • LAUNCHHELM_ASSISTANT_API_KEY: the provider's key, a secret. anthropic needs one; with the OpenAI ones, a server of your own may take none.
  • LAUNCHHELM_ASSISTANT_MODEL: the model. The OpenAI ones require it, as no model is offered everywhere.

LAUNCHHELM_ASSISTANT_EFFORT sets how much the model thinks per step; with the OpenAI ones it is sent as the reasoning effort only when set. The model must be able to call tools, as everything the assistant looks up goes through them.

With openai, nothing is stored on the provider's side: each request carries the conversation, and a reasoning model's reasoning comes back encrypted, so the model keeps it from one tool call to the next. A model or server that won't encrypt it is asked once without, and its reasoning is then not kept. The Chat Completions API returns no reasoning to keep.

Changing the provider or the model keeps every conversation. One continued under another provider is rebuilt from what the thread shows (questions, answers, and what the assistant looked up), without the earlier model's reasoning.

LAUNCHHELM_ASSISTANT_WEB_SEARCH turns on web search, and LAUNCHHELM_ASSISTANT_RESEARCH_URL with LAUNCHHELM_ASSISTANT_RESEARCH_TOOLS connects an MCP research server with an allowlist of its tools; list only tools that read. Both run on the provider's side and need anthropic. The README's configuration table lists every setting.