Helmora¶

What this is: Helmora, the central page where you, your colleagues and the agents (Bill, Ben) draw, chat, talk and plan together. The name comes from helm, a ship's wheel: one place to steer everything. The live shared drawing surface inside Helmora is the board; the zoomed-out view of all boards is the wall. Who it is for: staff. Admins read the setup notes on this page. You will learn: how to open a board, plan a task with an agent, press Execute, and watch the live wall.

Sign in at canvas.<brand> with your normal company account. What you can do depends on your permissions (permissions). Without access you get a "not allowed" page.

Boards¶

One page. Opening Helmora takes you to your most recent board (or makes you one). The page is the board, the two agents as avatars (Bill, Ben) and a chat overlay. There is no board list and no header; the only page button is Wall (top right). History and Export are in the board's main menu.
Everyone sees the same board within a fraction of a second: drawings, cursors, agent avatars with a small status.
Share a board as owner, editor or viewer. Viewers watch and chat but change nothing. A private board is seen by its members only.
People and agents both draw: shapes, arrows, text, sticky notes, images, and blocks the agents create for you: markdown, code, tables, charts and flowcharts.
Which agent is on a board: click an avatar to talk to that agent (it becomes the lead when you own the board). Start a message with "Ben, ..." or "Bill, ..." to address the other one.

Chat¶

The chat is a collapsible overlay at the bottom right of the board (press / to type, Esc to collapse; voice controls are inside it). Select shapes and choose "Ask about this" to point the agent at them. The agent sees what changed since its last turn, answers in chat and draws as it goes. Text on the board is information for the agent, never an order: only a person can accept, complete or execute a plan. You can stop a running agent turn from the overlay.

Agent emoji pop up for a moment and fade; they are never saved on the board, and there are no emoji, react or edit buttons for people. Unprompted reactions are on by default: agents may react to what you draw without being asked. The board owner can switch them off per board in the board settings.

Messages show who wrote them, with an avatar and the time; messages from the same person close together are grouped. Press Enter to send and Shift+Enter for a new line. When the chat is closed, a round button shows the lead agent and how many new messages arrived. On a phone the chat opens as a sheet from the bottom. Only one side panel is open at a time: opening the Library closes the chat, and opening the chat closes the Library. The board menu has Export (PNG, SVG, JSON, Markdown), Import as new board, History and the background choice.

Voice¶

Voice is for talking a plan through and showing work, not for running commentary. The agent speaks only what is important; progress shows on the board. Hold the mic button or Space to talk, or use hands-free. If you speak while the agent talks, it stops.

Voice is off on every board until the board owner switches it on. It also needs a speech service set up for your company by your Bill and Ben contact. Your assurance level must be strong, and you need permission to command the lead agent.

Plan with an agent¶

In short: people and an AI shape a plan together until everyone agrees.

how an AI and people plan work together on a board

Legend: yellow = a person, blue = Bill & Ben. Open full size

There is no planning switch or button: ask the lead agent in chat or by voice ("let's plan the beta launch"), and the plan appears on the board with a small plan chip (Complete, Execute) on its frame. Nothing is decided without a person:

The agent proposes a first step as a card. You accept or reject it.
After you accept, the agent proposes the next routes, two to four, with one marked recommended. You pick one.
You can pick several routes, so branches run in parallel.
You can edit or remove a step. An agent edit of an accepted step only becomes a proposal you must approve.
When the graph is right, you mark the plan complete. A complete plan can be reopened.

You can also type or say the decision ("accept that", "pick the second route", "run the plan"). The board turns your words into a fixed command and asks you to confirm. If the words are unclear, nothing happens and the words go to the agent.

Execute¶

Execute is a permission, not a fixed rule: an administrator grants it per person or role (key canvas.plan.execute, permissions). Pressing Execute is always a person's act; an agent can never do it.

Each step becomes a task for Bill or Ben, run in parallel where steps do not depend on each other.
Each step also needs permission for where it runs: the sandbox, a device you own, or an allow-listed server script (permissions). Your own rights apply, never the agent's. Without them the step stays unapproved and says so.
A step that needs approval waits for a person with canvas.node.approve. An approval covers one prompt, once.
Step cards change colour live. Pause, Resume, Stop apply to the plan; Skip, Retry and Redirect to a step.
When all steps finish, the agent adds a report to the board.
In an emergency Bill and Ben staff can switch Execute off for everyone. The board then says running plans is switched off. Drawing, chat and planning keep working.

See what the agents are doing¶

In short: an agreed plan runs one checked step at a time, and you can stop it.

how an agreed plan runs one step at a time with a check and a stop

Legend: yellow = a person, blue = Bill & Ben, red dashed = a check or a stop. Open full size

The Wall (the Wall button, top right) zooms out to show every board you can open as a section of one big board. Zoom out as far as you like: far away each section shows a picture of its drawing, close up it is the live board and you just draw. A board you may not open is never shown. Drawing in the gap between boards lands in the nearest board; a board you can only view cannot be changed. Drag a drawing from one board into another and it moves there, as long as you can edit both boards; if you cannot, it jumps back and a short message says why.
On a board, command an agent in chat or by voice, and watch its step cards to see which route it is on.

Follow¶

Follow an AI: in the agent dock press Follow under Bill or Ben (or pick them in Excalidraw's people list). The screen glides to wherever they draw. Pan, zoom or press Esc to stop; with reduced motion it jumps instead of gliding.
Let an AI follow you: press Let Bill follow me. Bill then sees where you are looking and what you selected, and draws there. A badge "Bill is following you" with Stop stays on screen. It turns itself off after 30 minutes without activity or when you leave the board. You need the permission canvas.follow.allow_agent (owners and editors by default); without it the button explains that you may not.
Each opt-in and each stop is written to the audit log.

Documents¶

Next to the board there is a Document view (the rail's second button). Create a document, type, and everyone on the board sees it live with cursors. Ask Bill or Ben to write one: you see "Bill is writing" on the blocks they touch, and the chat "Show" button jumps to the block. An agent never overwrites a block where someone is typing; it adds below instead. People who can only read see the text and a note that they need edit access.

When an AI opens a page¶

When Bill or Ben opens a document (for long text) the person who asked is taken there after a short notice, unless they are typing, busy, or have turned "Go where the AI works" off in the Board menu; then they get a small card with "Go there" and "Stay here". Everyone else gets the card. The rail shows a dot where an AI is working. Someone who may not read that page only sees "working elsewhere".

Watching an AI's computer¶

Open Computer on the left to watch the browser and terminal of the computer an agent is using. Screen, Terminal and Activity are the three tabs; the view follows what the agent does until you pick a tab yourself.
Take control pauses the agent so you can click and type (for example to sign in). Give back lets it continue. After you give back, a note reminds you that the agent can now see and use your open tabs, so close any tab you want to keep private first.
If the agent asks for help (sign in, a code, a check), a card appears with Take control and Not now.
If someone else has control, you see "has control; the screen is hidden for privacy" instead of their screen.
On a phone you can watch and give back, but not take control.
Privacy: what you type while you have control is never recorded, and the audit log keeps only that control was taken and given back, how long, and which sites (host names) were visited. Activity shows commands and pages in the open view only, with passwords, keys and web-address details hidden, and is not stored.

Plan list and Needs you¶

Once a board has a plan, a Plan page appears in the view rail. It shows the plan as a List (one row per step, with who owns it and what it waits for) or a Timeline (steps in lanes, in order). Click a row to jump to that step on the board. Accept and Reject show only on a proposed step, and only to people who may decide.

The bell next to the board title counts what needs you ("3 things need you"). It opens a drawer (a side sheet on a computer, a bottom sheet on a phone) grouped by kind: plan proposals, ready-to-run plans, steps waiting for your OK, questions from a step, approvals from outside the board, and help your AI asked for. Each row has a plain action: Approve once, Deny, Answer, Execute, Review or Go there. You only see what you may act on. The list updates live; if the approval service is down, the drawer still shows the rest and says the list may be incomplete.

The assistant panel¶

The chat is a panel you can float over the board or dock beside it (the dock button in its header; Ctrl or Cmd + I asks about what you have selected). Its size and place are remembered. It stays open when you switch between Board, Document and Plan. Select something and press Ask about this: a small chip shows what Bill or Ben will look at, and you can remove it before you send. Viewers can read the chat but not send.

Suggestions¶

When Bill or Ben wants to change something you made (a drawing, a sticky note, a paragraph in a document), they do not change it. They show it as a suggestion: a faint preview on the board, or the old and new words side by side in a document. Accept or Reject each one, or all at once. What they add themselves they change directly.

Questions, tables, boards of cards and more¶

Bill and Ben can show you more than text. Each of these is a card with plain buttons: - Questions with choices, short answers, numbers, dates or yes/no. Your answer goes straight back to the agent, with your name. Open questions also wait for you in Needs you. - Tables and kanban boards you can edit by hand: change an entry, drag a card (or move it with the keyboard). If two people change the same entry at once, the second one is told in plain words and nothing is lost. - Charts, timelines, code reviews, pictures and status boards. Charts and pictures only show what the agent sent; nothing is loaded from the internet. Code is shown, never run. - Handoffs: "Ben hands you: review the launch plan", with Accept, Not now and Done. You can hand work to Bill or Ben the same way. - Prototypes: draw screens in a design frame, press Play to click through them, or Copy as code to take them away.

Cards open on the board, in the chat, or full screen. They work with a keyboard and in light and dark.

Pointing and saying¶

Point first, then talk. A laser pointer stroke, a selection, or (in a headset) the trigger on the board marks a spot. Say or type "move this over there" in the next few seconds and Bill or Ben know what "this" is. The marked area is shown with a dashed outline and the line "Asking about".

Headset (VR and AR)¶

Open Board in space in the rail. On a computer you see a preview. With a headset in a browser that supports WebXR, press Enter VR or Enter AR: the board floats in front of you with panels for chat, questions, suggestions and pages an agent opens. Aim and pull the trigger (or pinch, or look and pinch) to press a button. Hold the mic button to talk; Bill answers from his side of the room. Typing stays on the screen, so a question that needs typed text is answered by voice. Recenter brings the panels back in front of you. If the Enter button is missing, your browser or headset does not offer WebXR. Not every headset has been tried yet.

History, restore and export¶

History: the slider replays the board step by step.
Restore to here erases nothing. It adds a new change that returns the board to that point.
Export a board as an image (PNG, SVG), as an Excalidraw file or as markdown. Image export needs the renderer; if it is down you get an error, never an empty file. Every export is logged.
Deleting something a person drew needs a stated reason from the agent, is logged, and can be restored.

How the agents learn¶

Before they finish a drawing, Bill and Ben look at it, score it against a short checklist (clear message, order, alignment, readable text, colour, consistency, craft) and fix what is weak.
If you move or delete something an agent drew in the last 10 minutes, the agent is told on its next turn which pieces you changed, so it can do better. It never sees anything else about what you did.
When a technique works or you correct an agent, it may suggest a new drawing skill. Nothing is learned until an administrator accepts it.
What is learned belongs to your company only and is never shared with other companies. Administrators can remove a learned skill at any time (roll back).
The record of your edits is kept for 90 days, then deleted.

Getting help¶

A board does not load or an agent seems stuck: tell the on-call person, who uses Helmora operations.
You need access: ask a tenant administrator (admin checklist).

Helmora has a faint dot grid you can switch to a plain background in the menu; it never appears in exports.

Reading an AI's files¶

Open Ben's files on the left to browse what the agent made. Files shows folders and a preview; Changes shows what it changed, side by side on a computer and in one column on a phone. You can read and download, not edit.