Work and runs

Day-to-day work in a product happens in work items: conversations, build runs, and design work. All three start from the same place, share one composer, and stream their runs inline. The co-founder is separate — one durable conversation per product, pinned in the navigation.

Start work (⌘N)

Press ⌘N or click "Start work" at the top of the sidebar's Work section. The view asks "What are we doing in {product}?" and offers mode cards:

Build is selected by default. The Co-founder card appears only for products in the Building state; a product still on the journey (or stopped) skips the cards — every start goes to the co-founder. Nothing persists until you send the first message: files stage locally, and the work item is created only on send.

The three kinds differ most in what they may touch:

Folders and archiving

Work items list in the sidebar under "Work". Each row's context menu offers Rename, Move to folder, Archive/Unarchive, and Delete; dialogs use the item's own noun ("Delete this conversation?", "Rename build run"). Deleting removes the item and its messages permanently but never touches repository files. Move, archive, and delete are disabled while an item has a live run.

Create folders with the "New folder" button, and start work directly inside one with "Start work here". Deleting a folder keeps its contents: "Work inside it is kept and moves back to the top level." Archived items hide behind the footer's "Show archived (n)" toggle.

The composer

Enter sends, Shift+Enter inserts a newline, and the Enter that commits an IME composition never sends. Run configuration is edited in the pill cluster and persists to the work item immediately.

Model

The model picker lists only models from providers you have configured, grouped per provider. Vision-capable models carry an image icon ("Understands images"), fast-capable models a lightning icon ("Supports fast mode"), and hidden models sit under "More". Type into the search field ("Search or type a model id…") and any id not in the catalog appears under "Use typed model" as Use "{id}" with {Provider}, so new models work the day they ship. See Providers and models.

Effort and speed

The effort pill ("Reasoning effort") offers "Default effort" plus the provider's own levels:

Fast-capable models add a separate "Fast" pill — a fast service tier, not another reasoning level.

Roles

Build runs carry a role pill ("Agent role"): Coder (the default), Planner, Heavy Planner, Reviewer, or System Architect. Chat pins the Assistant role; design work shows a persona pill instead (Auto, Specialist, Designer); co-founder conversations show neither — they always present as flowless, with model and effort still first-class. Build runs also get the Single agent / Flow toggle; see Flows.

Files and artifacts

Attach up to 10 items per message (files and quotes share the cap), 10 MB per file: source and text files, CSV, JSON, logs, PNG, JPEG, GIF, WebP, and PDF. Files stage as removable chips and upload only at send; invalid files are refused with the reason in red. The pen-nib pill ("Attach product artifacts") is different — attached artifacts persist on the work item and are injected into every one of its runs. See Workspace and artifacts.

Images and vision

Paste images straight from the clipboard. Vision is gated per model: pasting with a model that can't see images is refused, and images already staged when you switch to a non-vision model are dropped at send with the warning "Attached images will be dropped — {model} can't see images." and their chips marked amber. Cerebras and Grok endpoints accept PNG and JPEG only — other formats are stripped for those providers even on vision-capable models.

Dictation

The mic button ("Dictate") transcribes speech with a local Whisper model — fully on-device, Metal-accelerated. If no model is installed, pressing it opens Settings → Voice instead.

1. Download a model once in Settings → Voice. Downloads are SHA-256-verified, and a finished download auto-selects that model.
2. Click the mic and speak. A timer and input-level bar track the recording, which caps at 120 seconds. Esc cancels and discards the audio.
3. Click again to stop and transcribe. The text lands at the caret, ready to edit before sending.

Quoting the transcript

Select text inside a single message bubble and a small menu offers "Add to composer" and "Copy". Quotes stage as chips, count against the 10-item cap, and are truncated at 8,000 characters. They travel inline with your message — never as file uploads — and the model receives them as a "# Quoted from the conversation" blockquote section.

Quick-answer cards

When a top-level Chat or Co-founder reply asks you something, Motriz can render the questions as tappable answer chips under the message — at most 3 questions with 2 to 5 concrete options each, plus a free-text field ("Or answer in your own words…"). Submitting sends one composed message that reads like typed text. A card is answerable only while its message is the last in the conversation. Only top-level Chat and Co-founder runs get this contract — a delegated sub-agent can never address you directly.

Follow the run

Runs stream inline in the conversation — there is no separate run view. Tool calls, thinking, flow steps, sub-agents, and deep research all render as the same kind of row, and the final answer streams as a normal chat bubble.

Run details fold by default: the turn row carries a live activity label, a duration ticking once per second, and running totals in the form 41.2k in · 3.1k out · $0.86 · 3m 12s. Unfold it whenever you want the receipts — tool rows lead with an uppercase verb (Read, Write, Edit, Grep, Glob, Bash, Search, Fetch, Research, Publish, Delegate, Plan), and a "Thought" row shows the model's current reasoning, replaced as it moves on. A finished turn settles into a one-line receipt that unfolds back into exactly the rows it streamed. Delegated child runs collapse by default under a delegated badge. Token and cost usage rolls up the run tree and feeds the Dashboard's 7-day spend.

Stopping a run

Click the send button — now Stop — to end a run; the turn records "stopped" or "stopped after {duration}". While anything runs, the item's sidebar title shimmers and an orange "{n} RUNNING" pill sits on the Build row. Closing the window with the red traffic light only hides Motriz to the Dock — runs keep going. Quitting from the sidebar's "Quit motriz" button warns you first: "A run is still in progress — quitting stops it."

Automatic naming

Work items name themselves using on-device Apple Intelligence: a concise 3–8 word title, generated at most once per item — Motriz tries at run start and again after the run ends, and the first title that lands is final. A manual rename is never overridden — click the title in the header to set your own. Without Apple Intelligence there is no cloud fallback; titles simply stay "New work".

Deep research

Every run's toolkit includes deep_research: it delegates a question to a dedicated deep-research agent that "plans its own searches, reads sources, verifies claims, and returns a cited report with explicit coverage gaps and an evidence block." The researcher always runs on the calling agent's own provider and model.

Inside flows and co-founder conversations it runs as a real child run — visible, cancellable, and usage-tracked, labeled "Deep research: {question}" with the delegated badge and an expandable rail of its searches and reads. In Chat and design work the same researcher runs in-process, its tool calls and usage streaming into the calling step. Searches use Motriz's embedded web search; Claude Code and Codex bring their own native web tools for point lookups, but deep research is a Motriz capability and is served to them as well.