Chatting with CSuite.

1

~/Launch+

Grinder launch assets14 messagesBurr geometry research6 messagesQ3 report questions9 messages

23

Grinder launch assets⧉Gemma 4 12B ▾

4

Make me a hero image of the grinder on wet slate, hard side light, 16:9.

✓Thinking✓Generating imageGenerated image

Here is the hero shot: brushed titanium on wet slate, hard light from the left. Want a 1:1 crop for the product grid?2 min ago · gemma4:12b⧉ ↻ 🗑tool image_generate

Jump to latest ↓

567

Message…Add imageAdd documentWebSend

Open Chat from the rail, the middle entry of the second group. It is one conversational workspace where you talk to a model and make media in the same thread: ask a question, search the web, read a document you attach, generate or edit an image, a clip or a voice, or run a saved workflow. Everything it makes lands in your project folder as a real file.

1 · Conversations

Your project folder pill, the + for a new conversation, and every conversation with its message count. Rename or delete from a row’s More actions.

2 · Title

Names itself after the first exchange; edit it in place and your name wins.

3 · Models

The chat model by name, opening the picker and the More models slots the assistant generates with. Beside it, copy the whole conversation as Markdown.

4 · Thread

Your messages in bubbles on the right, replies as plain text on the left, with a step list and the media each turn produced. A Jump to latest pill floats when you have scrolled up.

5 · Composer

Message…, with Enter to send and Shift+Enter for a new line. Stays editable while a reply runs; Send becomes Stop.

6 · Attach and Web

Add image, Add document, and the Web toggle that lets the model search and read pages from your machine.

7 · Send

Or Resend when you are editing your last message.

Deleting a conversation stops anything it is still doing and removes it from your disk; there is no archive. History is stored locally in the app’s database and never leaves the machine on its own.

The Models pill in the header opens the chat slot and, under More models, the slots the assistant reaches for: Image, Video, Speech, Music, Sound effects and Transcribe. The chat model also writes the replies and the prompts it sends to those models, so there is no separate text slot. Any slot can be Auto. Ask for a kind of media whose slot is empty and the assistant stops, says so (“No image model is selected yet. I’ve opened Models for you…”) and opens the picker on that slot.

Rules you write under Account → Instructions (Global, plus the Chat section) are appended to the assistant’s system prompt on every turn.

• Streaming: cloud and local replies arrive token by token. The composer stays editable meanwhile.
• Reasoning: a thinking model’s thoughts show in a collapsible block, open while it streams and kept on the saved reply.
• Steps: a list ticks off each model round (“Thinking”) and each tool call (“Generating image”, “Searching the web for …”, “Cropping image to 1:1”) with a done or failed mark, so a long turn is never a blank spinner. A turn chains up to six tool rounds, then stops and says so.
• Footer: every reply names the model that wrote it; hover for the tool it used, the tokens in and out, and how long it took. Notices (“Stopped.”, a pick-a-model note) render muted and are not replayed to the model later.
• Context: history is budgeted from the model’s real window, oldest messages dropped first; your new message always goes in full. A reply that runs out of window says so.

• Regenerate the newest reply: the whole turn re-runs, tool rounds included, reusing the images you had attached.
• Edit and resend your newest message: the composer shows “Editing last message — Esc to cancel”, and on Resend everything after it is removed and the corrected version runs fresh, with its attachments.
• Stop genuinely cancels: the provider call is aborted and a running edit is killed. What earlier rounds produced stays in the thread, and a later round failing never takes an earlier image with it.
• Copy any message, or the whole conversation as Markdown from the header. Delete removes one message after a confirm.
• Keep up: the thread follows the reply until you scroll up; Jump to latest brings you back.

The speaker button under a reply (Read aloud) voices it with your Speech model, the markdown stripped to plain text, and plays it in a small player under the reply. Afterwards the button plays and pauses the saved clip. The clip is kept with the conversation rather than in your project folder. With no speech model set the button says so with an Open Models shortcut.

Ask in plain language and name the kind of thing you want: “make me a hero image of…”, “a 5-second clip of…”, “say this in a warm voice”, “a 30-second lo-fi bed”, “a door creak”. Before each turn the chat model reads your message, in any language, and decides which tools to offer: a kind is offered when you ask for it, follow up on it (“make it blue” after an image) or say yes to a reply that suggested it. A request naming no kind (“generate a cat”) gets one short question back instead of a paid guess.

• Images go to the Image slot; video to the Video slot; speech, music and sound effects each to their own audio slot, so a voice-over never runs your music model. Speech is only offered when you ask for it in words (say, speak, narrate, read aloud, voice).
• Settings ride in the sentence: “16:9”, “square”, “portrait”; for video “8 seconds”, “1080p”, “with audio” or “silent”; for music a length and a format. They are lifted out before the prompt is sent and checked against the model.
• Under Auto on a slot, a hint such as logo, icon, vector, photo or cinematic steers which model is picked for that call.
• Results appear in the thread with a caption (“Generated image”, “Generated video”) and the model’s own narration, and are saved into your project folder. Generated speech is kept with the chat instead.

~/Launch+

Grinder launch assets14 messagesBurr geometry research6 messagesQ3 report questions9 messages

Grinder launch assets⧉Gemma 4 12B ▾

crop it to 1:1

✓Cropping image to 1:1Cropped image to 1:1No model round: an obvious edit is matched before the model is called.

make the slate matte black and keep everything else

✓Thinking✓Updating imageUpdated image

Done: the slate is matte black now, the tamper and beans unchanged.2 min ago · claude-sonnet-5⧉ ↻ 🗑tool image_update

Message…Add imageAdd documentWebSend

Edits target the most recent file of that kind in the thread, so “it” means the last image, clip or recording. Two paths:

• Deterministic edits: crop or resize an image, crop, resize or convert a video, trim or convert audio. An obvious English request (“crop it to 1:1”, “resize the video to 720p”, “convert to WebM”) is recognised before the model is called and runs on the local engine, on any provider, with no token spent. A question never triggers one; a crop needs a standalone ratio such as 9:16.
• Creative revisions: “make the slate matte black”, “same scene at dusk” run image-to-image on the last image (“Updated image”) through your Image model.
• Validation: an argument the model gets wrong (an aspect ratio the file cannot take, a height off the ladder) is a tool error it can correct, never a silent default.

~/Launch+

Grinder launch assets14 messagesBurr geometry research6 messagesQ3 report questions9 messages

Grinder launch assets⧉Gemma 4 12B ▾

Make a video that starts on the first and ends on the second, 8 seconds.

✓Thinking✓Generating video▶Generated video · start and end frames from your two images

Message…Add imageAdd documentWebSend

A video request uses the turn’s own images as frames: one image attached to the message or made earlier in the turn becomes the start frame, two become start and end. With none, the two most recent images in the conversation are used. So “animate that” after an image, or “a video from the first to the second” with two attachments, both work. The Video slot’s model must take a start frame for this; a text-only model ignores them.

~/Launch+

Grinder launch assets14 messagesBurr geometry research6 messagesQ3 report questions9 messages

Grinder launch assets⧉Gemma 4 12B ▾

Run my podcast intro workflow.

✓Thinking✓Running workflow "Podcast intro"Workflow output 1▶intro-voice.mp3Workflow output 2

Ran workflow “Podcast intro” — 4 nodes completed. The cover and the voiced intro are above and saved in your project.2 min ago · gemma4:12b⧉ ↻ 🗑tool workflow_run

Message…Add imageAdd documentWebSend

Ask for one of your saved workflows by name (“run my podcast intro workflow”) and it runs headlessly with the same engine as the Workflows screen. Each File node’s media output is posted into the thread as an attachment, and the caption counts the nodes (“Ran workflow “Podcast intro” — 4 nodes completed”, with failures counted when some stop). A name that does not match lists the workflows you have.

~/Launch+

Grinder launch assets14 messagesBurr geometry research6 messagesQ3 report questions9 messages

Grinder launch assets⧉Gemma 4 12B ▾

The current chat model doesn’t accept imagesSwitch to a vision-capable model to attach them.

Drop files to attach…

✕✕MDbrand-voice.md ✕Which of these two labels reads better from a metre away?Add imageAdd documentWebSend

Attach any number of PNG, JPG, GIF or WebP images with Add image, by dragging them onto the thread, or by pasting from the clipboard. On a vision-capable model each is sent as real image input, so you can ask what is in them or compare two. A model that cannot see images says so with a toast (“The current chat model doesn’t accept images — switch to a vision-capable model to attach them.”) rather than ignoring them. The images you attach are also what the next edit, update or video uses first, before anything older in the thread.

• The Web toggle in the composer is on by default and offers search and page reading on every turn while it is on. Off, the chat is fully offline (“Click to keep this chat offline”).
• No key, no service: the search runs from your machine. A hidden, sandboxed window loads a search engine’s own results page (Google, then DuckDuckGo, Bing and Brave if one gives nothing), and the top four results are read and reduced to text. The first search of a session takes about ten seconds, later ones about four.
• Sources: pages are numbered per turn, the reply cites them as [1], [2] after the sentence they support, and the sources show as chips under the reply that open in your browser. Web text is framed to the model as source material, not instructions, and the reply knows today’s date.
• Budget: a quarter of the model’s window is reserved for web results, so a small local model reads less of each page than a large cloud one.
• Guard: addresses on your local network and private ranges are refused, because the URL is chosen by a model that has just read untrusted text.

Conversations, messages and attachments live in the app’s database on your disk. With a local Ollama model, the replies, the tool decisions and every deterministic edit run on the machine; with a cloud model your messages, attached text and images go to that provider under your key. Generated media follows each slot’s model. Web search runs locally either way, and turning Web off keeps a thread entirely offline. The full picture is in Privacy & data.

⌘ is Ctrl on Windows and Linux. The main guide lists every workspace’s shortcuts together.

• “No chat model selected. Open Models and pick a Chat provider.” Set the chat slot; Auto needs at least one tool-capable provider.
• “This chat model can’t make tool calls.” The picked model has no function calling; choose one the picker lists. The model still works in the Text workspace.
• My favourite model is not in the picker. Replicate and Hugging Face models have no tool surface, and a few Runware tiers cannot finish a tool call there (Gemini works through OpenRouter). Use them in Text instead.
• “No image model is selected yet. I’ve opened Models for you…” The slot for that kind is empty; pick one and ask again.
• It asked “what kind of thing?” instead of generating. The request named no kind; say image, video, voice, music or effect, or answer yes to its offer.
• “(Stopped after reaching the tool-call limit.)” Six tool rounds per message; split the job across two messages.
• “The current chat model doesn’t accept images.” Switch to a vision-capable model to attach images; documents still work on any model.
• “n files skipped”. Only PNG, JPG, GIF and WebP images and PDF, TXT, MD and HTML documents attach. “Couldn’t be read from where it was dragged”: save the file to disk first.
• “Pages 1–30 of 300 seen”. The document is larger than the window; digest it, ask about a part, or use a larger model.
• “Ran out of context window”. A thinking model reasoned into the window, or the history is long. Start a new conversation or pick a model with a larger window.
• “Wait for the current reply to finish before attaching files.” Attach after the turn, or press Stop.
• Web search found nothing or was slow. The first search of a session takes about ten seconds; engines are tried in turn and a consent page skips Google for a while. A local address or private network is refused by design.
• “No workflow named …”. The reply lists the saved names; use one exactly, or save the workflow first.
• “No text-to-speech model is selected.” Read aloud needs the Speech slot; press Open Models.
• A crop did not happen locally. The deterministic matcher is English-only and needs a standalone ratio; in other languages the model runs the same edit tool.
• “Stopped.” with an image already in the thread. That is intended: earlier rounds persist the moment they finish, and Stop only cancels what was still running.