Using the AI assistant
One assistant that takes a goal, plans, builds with real modelling tools, measures what it made and repairs what came out wrong — into a scratch tab, so your document is untouched until you keep it.
The right-panel assistant does not draw pictures of parts — it drives the same commands you do, so what it makes is ordinary features in the tree that you can select, re-dimension and edit afterwards.
There is one mode. Ask it a question ("why is this feature red?", "what does this weigh?") and it answers, inspecting the model to do so. Give it a goal ("a 40×20×10 bracket with 2 mm fillets") and it plans, builds, measures what it made, and repairs what came out wrong.
Set-up
Open ⚙ and pick a model. Its provider lights up in the Providers row below, with just that one provider's setup open — so what you need next is in front of you, and the other seven stay out of your way.
The list only offers models that can read a drawing. That is not a preference: ⊕ Mark and an attached print are how you point the assistant at real work, and a model that cannot see an image fails those silently. Claude, GPT-4o/4.1, the Gemini family and the Chinese frontier vision models (Kimi, GLM-4.5V, Qwen-VL) all qualify.
- A local Ollama model needs no key and nothing leaves your machine — install a vision one (
llava,llama3.2-vision,qwen2.5vl), then click the Ollama chip and Detect installed. A text-only local model is found but not listed, and it says so rather than reporting nothing found. - Cloud models use your own API key. It is stored in this browser only, and requests go straight to that provider.
- Already using something else? A model you had selected before stays selected and keeps working — it appears under Current in the list.
The Providers row
One chip per provider. Click any chip to open its pane — you don't have to choose a model first, and only one pane is ever on screen. A chip is marked ✓ when that provider is ready to use, so hiding the panes never hides where you stand.
- A cloud chip opens one API key field. Ready means a key is stored. Forget keys clears them all.
- The Ollama chip opens the local setup instead — no key, the Detect installed button, and the server address if your Ollama runs on another machine or a non-default port. Ready means a usable model was detected.
A cloud provider only gets a chip if it currently offers a model you could pick, so you're never asked for a key you'd have no use for. Ollama always has one: its pane holds Detect, and hiding it when nothing is installed would hide the very button that fixes that.
Picking a model selects its provider's chip for you, so the thing you now need is already open.
Advanced holds what most people never need: a different model for the agent's Planner or Verifier (both default to the one you picked), and a custom API endpoint for the cloud provider in focus.
It also has "Also show models that can't read a drawing." Tick it and the text-only frontier models come back, labelled · text only so the difference stays visible — and a DeepSeek chip appears, since DeepSeek's hosted API has no multimodal model at all and is otherwise the one provider you'd never need a key for. Those models still build from a typed description ("a 40×20×10 bracket with a 6 mm bore"); they just can't read an attached print or a ⊕ marked face.
Letting Ollama answer this site
Ollama only accepts browser requests from localhost by default, so on a real domain — including hammer8.ai — it refuses the connection and no models are detected. This is Ollama's setting, not a hammer8 one: your browser reaches it fine, Ollama declines to answer.
Add this site to OLLAMA_ORIGINS, then restart Ollama (a running server never re-reads it):
macOS — the menu-bar app: `` launchctl setenv OLLAMA_ORIGINS "https://hammer8.ai,https://www.hammer8.ai" `` Quit Ollama from the menu bar and reopen it.
Windows — Settings ▸ System ▸ About ▸ Advanced system settings ▸ Environment Variables, add a user variable OLLAMA_ORIGINS with that same value, then quit Ollama from the tray and reopen it.
Linux (systemd): ``` sudo systemctl edit ollama
[Service]
Environment="OLLAMA_ORIGINS=https://hammer8.ai,https://www.hammer8.ai"
sudo systemctl daemon-reload && sudo systemctl restart ollama ```
Running ollama serve yourself — set it inline: `` OLLAMA_ORIGINS="https://hammer8.ai,https://www.hammer8.ai" ollama serve ``
Then reopen ⚙, click the Ollama chip and press Detect installed. OLLAMA_ORIGINS=* also works, but it lets any site you visit talk to your local models — name the origins instead.
A non-default port is fine (set it in that same pane, under Ollama server), but the server must stay on localhost/127.0.0.1: the browser blocks plain-http calls to any other host from an https page. Use 127.0.0.1, not [::1].
What it can reach
Around 120 tools: sketches with real constraints and dimensions, every solid feature, surfaces, sheet metal, patterns and mirrors, assembly mates, drawings, measurement and mass properties, materials — and the Help Center itself, so it can quote these articles back at you.
It is also reading your model: the feature tree, what you have selected, and real measured values. Answers are about your part, not a guess from the wording of your question.
Giving it context
- ⊕ Mark a face, edge, vertex or feature — and in an assembly a component or one of its faces, on a drawing a view — to attach it before you ask. The chip carries the entity's real id, and the assistant is told to act on exactly that, so "fillet this R3" lands where you pointed instead of somewhere plausible. Pick several; Esc or a click on empty space finishes.
- Attach an image to a message — on a vision-capable model it can look at a sketch or a reference photo.
How much rope to give it
Set the autonomy on the toolbar:
- Auto — plan and build.
- Confirm — show the plan and wait for your approval.
- Plan only — never executes; useful for a second opinion on approach.
The build goes into a scratch tab, so your document is untouched until you press Keep this build; Discard closes the tab and takes the attempt with it. On a vision-capable model the agent also looks at the rendered viewport to check its own result.
Design memory
Memory travels inside the file, so a part carries its own intent to every later session and to whoever opens it next.
/remember wall thickness is 2 mm— one fact./capture <a design brief>— paste a brief and let it extract the facts./init— seed memory from the model that is already built.
Facts are grouped as constraints, decisions, intent, conventions and notes, and the agent reads them before it builds.
When it fails
- Nothing happens when you send a message — no model is configured; the ⚙ popup opens by itself in that case.
- The model chats but never builds — it is not tool-capable, or it is a very small local model. Try a larger one; tool-calling is what turns a reply into geometry.
- The agent's build is wrong — Discard and restate the goal with the numbers explicit. Nothing reached your document.
Related: Your first part (sketch → extrude) · Equations & global variables · Fix a rebuild error