Use the chat builder

The AI Chat Builder lets you describe a workflow in plain English. It opens as a floating, draggable, resizable window over the canvas.

Opening it

• Ctrl/Cmd+L toggles it open/closed.
• Click the AI Builder button in the canvas controls (purple gradient button with a star icon).

What it does

You type a message describing what you want, "Build a 3-agent research → writing → publishing pipeline" or "Add a reviewer that critiques the writer's output", and the chat builder applies structured edits to your workflow.

Behind the scenes, it emits typed operations (add agent, connect, set instructions, etc.) which Cerevisor applies the same way you'd apply them by clicking. Every edit goes through the same undo stack as your manual edits; Ctrl+Z works.

The window

Position

Drag the header bar of the window (not the buttons) to reposition. Position is clamped to the viewport: the window can't disappear off-screen.

Resize

Drag the bottom-right corner handle (cursor becomes nwse-resize). Constraints:

• Minimum: 320 × 320 px.
• Maximum: 960 × 1000 px.
• Further clamped to your current Cerevisor window size minus a small margin.

Minimize

Click the minimize button in the header. The window collapses to a 240 × 44 strip. Click again to restore to your last size.

Reset

There's no dedicated reset button. If the window ends up in a weird state, restarting Cerevisor re-clamps it.

The chat input

A standard textarea with these behaviors:

• Enter submits.
• Shift+Enter inserts a newline.
• The placeholder text suggests what to try.
• The input grows vertically as you type, up to a cap.
• Typing mid-turn queues your message. If you send while the builder is still working, the message isn't bounced — it's queued and fires as the next turn once the current one finishes.

While the builder works, a single activity line above the input shows what's happening right now — one place to look, so if nothing is showing, nothing is running. A Stop generating button appears whenever there's something to stop.

Attachments

Files

The paperclip button opens a file picker — markdown, PDF, text, image, code. You can also drag files straight onto the chat window, or paste an image from the clipboard (e.g. a screenshot) directly into the input. The chat builder reads the content as part of your message.

For images: included as an image the model can see (a warning appears if the current model can't process images). For PDFs: text-extracted. For markdown/code/text: included as marked attachment text so the model knows it's an attachment, not your prose.

Staged attachments show as chips above the input. Click the X on a chip to remove. You can send an attachment with no text at all ("here's a screenshot, what do you see?").

Skills

Drag a skill from the Skills panel onto the chat window to stage it. Staged skills appear as chips; when you submit, the chat builder considers them when proposing edits. Useful when you want it to design a workflow around specific skills you have.

Plan vs. Quick mode

A toggle in the window header:

Mode	Behavior
Plan	The builder asks a few clarifying questions first, then drafts the team.
Quick	Builds immediately from your first message.

Permission posture

A strip below the model strip shows the app's approval mode (Oversight / Auto / Hands-free) — the same switch that governs how much runs ask before acting. See Run and monitor for what each mode does.

Progressive mode

A per-workflow toggle in the window. When on, the builder doesn't just build: it builds the next step, runs it, and reports back — no Run click needed. Toggle off for the classic build-only chat.

Model strip

A strip under the header shows which model the chat builder is using and where that choice comes from:

• from workflow override: this workflow has a specific model set for the chat builder.
• from global override: the chat-builder config has a model set.
• Provider default: using whatever model the provider defaults to.

Click it to open a two-column picker: providers on the left, that provider's models on the right. Pick a model to pin it, use Use default for the provider's default, or Reset to default to clear the override. Per-workflow overrides live in Settings.

Run report

When a workflow run completes, the chat panel auto-appends a markdown Run report message showing every terminal agent's output. This makes the chat window double as a "what did my workflow produce?" log.

The report:

• Concatenates every leaf agent's output under per-agent headers.
• Truncates at ~12,000 characters with a footer pointing back to the agent card.

Stalled detection

If the chat builder's generation stalls (no progress for several seconds), a "Generation stalled, Rephrase?" chip appears below the message. Click it to focus the input and try a different wording.

Workflow-incomplete hint

If your workflow has obvious gaps (no agents, no input configured, missing required fields), a "Workflow incomplete" chip appears above the input. Click it to jump to the relevant settings.

What the chat builder is good at

• Sketching the rough shape of a workflow from a description.
• Adding agents you'd otherwise add by hand.
• Connecting agents you'd otherwise wire manually.
• Filling in role descriptions and instructions.
• Quick edits ("rename the second agent to 'Senior reviewer'").

What it's not good at

• Long, multi-step refactors. Do them in pieces.
• Setting up MCP servers, providers, or other configuration outside the workflow.
• Deciding what your workflow should do. It builds the structure you describe; it doesn't pick the goal.

Provider caveat

The chat builder needs a chat API. Codex CLI and Cursor Agent providers can't drive the chat builder; they have no chat-only mode. If your default provider is one of those, configure a different provider for the chat builder under Settings → Chat Builder.