The Panel
Heimdall sheet beside every page, end to end: opening it, the context chip and the eye, the start screen, threads, the composer, the model tiers and the review switch, the work timeline and the streamed answer, buttons and follow-ups, confirmation cards and the budget line.
Heimdall panel is a sheet that slides in from the right of any page. The page keeps its width and stays readable behind it, and because the panel is mounted once in the layout it keeps your thread open while you move between pages.
Your first turn
Open it. Press Ask Heimdall in the sidebar, under the search, or press Ctrl+/ (⌘+/ on a Mac). The shortcut is printed beside the sidebar entry, the way the search shows its own.
Check the chip. Under the header a chip names what Heimdall is looking at: the project whose page you are on, or Overview when you are on none. That context travels with every message.
Ask. Type in the composer and press Enter, or tap one of the seven questions the start screen offers. Shift+Enter breaks a line. The small Auto button beside Send picks the model tier and switches review on.
Watch it work. A timeline grows between your message and the answer: the thinking as a faded line, every tool as a step with its arguments and the time it took, "Working for 3s · 2 steps" counting above; then the answer arrives word by word, formatted. A step that failed says why under its arguments: the refusal a tool answered with, or that the run reached its step limit before the call, in which case Heimdall carries on and runs it in the next step.
Act on the answer. Press a button to open what it made or read, tap a follow-up to send the next question, or Confirm a card when it asks before changing something.
The panel, piece by piece
Header
Heimdall's mark and name, the Chat and Council tabs, then one group of buttons: the Threads button that lists your recent conversations with New thread at the top, the thread's menu while a thread is open, Widen (the panel grows to the sidebar's edge and the conversation keeps a reading column; Narrow brings it back, and the browser remembers the choice), and the close button. Esc closes it too.
Chat and Council
Two tabs beside the name. Chat is the conversation. Council shows the sittings of Heimdall's council on the project in view, the newest first, with a card per voice and the verdict pinned on top, and lets the people who may decide accept or dismiss each finding. The tab follows the project you move to, and a record's page opens it on the sitting it just convened.
Context chip and the eye
The project or page in view, and, while the eye beside the chip is lit, the text of the page itself, read at the moment you send: what is on screen first, the rest of the page after it, up to 8,000 characters. So "why is this one amber?" or "add up this column" works on any page, and "pause the endpoint I have open" needs no name. Tap the eye to keep Heimdall off a sensitive page; your browser remembers the choice. What it reads counts toward the turn's tokens like the rest of the prompt.
Start screen
An empty thread shows seven day-to-day questions, each a quick read that changes nothing: which projects are in development, what changed this week, who leads each project, your unread notifications, which endpoints are failing, who is on your team, which tokens you hold. Tap one and it is sent as if you had typed it.
Transcript
Your messages on the right, Heimdall's under its mark, each answer's work folded to one line above it ("Worked for 4.1s · 3 steps"), the buttons, the follow-ups and the cards. Under every answer a small line says what it used: the tier and the model that answered, the prompt tokens with the cached part, the output tokens, the time, and what the answer cost. A table in an answer is shown as a card per row.
Composer
One line that grows as you type, up to six, with the model button and the send button beside it. A message may be up to 4,000 characters, in any language.
Budget line
Under the composer: "AI budget this month: $12.40 of $100.00 (12%)", or "AI spend this month: $12.40, no cap set" for a team without a cap. At the cap it reads "This month's AI budget is spent", with a Raise it link for a person who may.
Keyboard
| Keys | Does |
|---|---|
| Ctrl+/ | Opens the panel, and closes it again |
| Esc | Closes the panel |
| Enter | Sends the message |
| Shift+Enter | Breaks a line inside the message |
Who sees the entry
Ask Heimdall is shown only to people who may use Heimdall: the team owner and any team member whose role carries Use AI. A member whose role lacks it sees "Heimdall is not in your role" instead of the composer, and nothing is metered (Permissions & Safety). A guest of one project never sees it.
Threads
The Threads button in the header lists your recent threads, newest first, each named after its first message, with New thread at the top. Open an old one and the whole conversation is there, with each answer's work timeline, its buttons and its cards. A thread nobody has used for ninety days is removed.
A thread opened from a project page stays about that project for its whole life; a thread opened elsewhere takes the project in view turn by turn. When no project is in view and the request needs one, Heimdall lists your projects and asks which, unless you have only one.
While a thread is open, the ⋯ menu beside Threads offers two more things:
| Action | What happens |
|---|---|
| Archive thread | The thread leaves the list and nothing continues it. It keeps its conversation and sits under Archived threads in the switcher, where one tap restores and opens it. |
| Delete thread permanently | After you confirm, the thread goes for good with its conversation, its tool runs and its cards. There is no undo. |
A message whose turn could not run, because the model was unreachable or the budget was spent, leaves no thread behind: the notice appears under the conversation, your words stay in the composer, and the switcher stays as it was.
Model and review
The Auto button beside Send opens the model menu. Auto, the default, lets Heimdall pick the model for each request: a small router model reads the message and sends a one-lookup question to the simple model and anything that writes, drafts, compares or takes several steps to the involved one. The three named tiers pin a model, and the dots beside each say what it costs relative to the others.
| Tier | Best for | Cost |
|---|---|---|
| Auto | Everyday use; the router picks per request | Standard cost |
| Lite | Quick reads and simple edits | Lowest cost |
| Pro | Sharper answers and multi-step work | Standard cost |
| Max | Deepest reasoning for the hardest requests | Highest cost |
Under the tiers, Review every change makes every tool that writes wait for a card, the additive ones included, so a new webhook endpoint is shown to you before it exists. With review off, only changes to what already exists wait. A thread remembers the tier and the switch you last used on it.
While it works
The work timeline
Between your message and the answer a timeline grows as the turn runs: the model's thinking as a faded line with a hollow dot, every tool as a step with the arguments it used hanging off it on a braid of lines and the time it took at the right, and "Working for 3s · 2 steps" counting above. A running step shows a hollow dot too. Under the header a line in Heimdall's own voice says which phase the turn is in: "Bifrost firing up…" while the turn is routed, "Researching the projects across the nine realms…" while it reads, "Forging an epic in Asgard…" while it writes, "Heimdall speaks…" while the answer streams.
Folded once it is done
When the answer is in, the work folds to one line, "Worked for 4.1s · 3 steps". Tap it to open the rail again, tap a thought to read it in full; a failed tool shows a red dot with "Failed" under it, a cancelled card a grey one.
The answer
Streamed word by word and rendered as it arrives: bold, lists and short tables read correctly even mid-sentence.
After the answer
Buttons to the page
When a turn made or read something that has a page, a purple button opens it: View CRM sync for the endpoint it just paused, Open the deliveries for the log it worked through, Open the Notifications Center after it marked everything read. The button navigates without leaving the thread.
Follow-ups
Under the latest answer, two or three secondary buttons with the requests you are likely to make next, phrased as you would type them ("Retry the dead deliveries", "Who leads it?"). Tap one to send it.
The page follows the writing
While Heimdall writes records the page follows it: when it creates epics the Epics page opens behind the panel and the table fills in row by row as each one lands, the sidebar's counts climb with them, and a band of light runs round the edge of the window until the turn ends. A turn that changed something else refreshes the page behind the panel, so an endpoint it paused shows as paused in the table beside you.
What it wrote
When a turn creates or changes a record, a card appears under the answer: the kind (an epic, a story, a client, a glossary term), its reference where it has one, its title and where it stands, with Open to the record's page. A turn that wrote more than a handful folds the records of one kind into one card ("Components · 5 new", the first three names and how many more) that opens the page they live on, and shows four cards with Show N more for the rest. The cards read from what the tools answered, never from the prose, so a record the model mentions but did not write gets no card. A turn that convenes the council shows the sitting itself in the thread as the conversation it is: each voice's research, reading, answers and refinements as they land, Heimdall's plan and verdict at the end, and each finding carries Accept and Dismiss for the people who may decide, the same as on the Council tab.
Start a project this way: tell Heimdall about the client and what they need, in a paragraph. It creates the client and the project, writes the components, environments, people, epics, stories, requirements, rules and glossary as drafts, asks only for what it cannot infer, and convenes the council on the result; the walkthrough is on starting a project in conversation.
Confirmation cards
Anything that changes or removes what already exists stops before it runs; creating something new does not, unless Review every change is on. An amber card shows Confirm before I do this, the operation in words, every argument it will use, and two buttons.
- List webhook endpoints0.3s
- search · CRM sync
- Pause webhook endpoint
- endpoint_id · CRM sync
- endpoint_id
- CRM sync
Confirm runs it and Heimdall finishes its answer; Cancel declines it, and nothing happens. A card expires after ten minutes; ask again and Heimdall prepares the action afresh. Confirming is a model call of its own, because Heimdall reads the result and writes its answer, so it is metered like a message; cancelling costs nothing.
One card, one action
Heimdall never splits a destructive action into smaller pieces to avoid the card, and never repeats an action you cancelled. A tool that refuses because your role does not allow it is reported plainly; it does not look for another way round.
The budget line, state by state
| The line reads | Meaning |
|---|---|
| AI budget this month: $12.40 of $100.00 (12%) | The team's month against its cap; every turn and every other AI feature moves it |
| AI spend this month: $12.40, no cap set | The team chose no cap; the spend is shown for the record |
| This month's AI budget is spent · Raise it | The composer refuses to send until the cap is raised or the month turns |
Usage & Budgets has the whole meter.
When something goes wrong
| You see | What it means |
|---|---|
| "Heimdall could not reach its model just now" | The provider did not answer; nothing was changed or metered. Try again in a moment. |
| "This answer failed before it finished. Nothing was filed for it." | The turn broke mid-way; the thread keeps the note and nothing is filed in the ledger |
| A red dot on a step of the work, "Failed" under it | That tool failed; Heimdall says what it could not do |
| "Heimdall is not in your role" | Your role lacks Use AI; ask the team owner to add it |
| "This month's AI budget is spent" | The cap is reached; raise it under Settings → AI or wait for the next month |
Questions
Related
- Usage & Budgets — the meter behind the budget line
- Permissions & Safety — who may use it and what the model is shown
- Settings → AI — the cap, the statement and the models
- MCP tools — every operation Heimdall can run, under its kind
How is this guide?
Last updated on
Heimdall
The operator inside Zuse Bifrost: ask it about your projects, teams, webhooks and history in plain words, have it do the routine work, and confirm anything that changes what already exists. Beside every page, metered against the team's monthly AI budget.
Start a Project in Conversation
Tell Heimdall about a client and what they need; it creates the client and the project, writes the record as drafts, asks only for what it cannot infer, and convenes the council on the result inside the conversation.