Docs menu · Using the AI

Documentation

Using the AI

The AI panel is the AI tab of the bottom dock. Type what you want, press Enter, and the AI builds a node graph through the same tools any script could use. Everything it does is visible: the nodes it adds flash in the graph, the viewport updates, and the AI log lists its plan, messages and every tool call.

Providers

Pick the provider in the first menu next to Send, and optionally a model in the second. Tessera remembers your choice. Tessera has no AI service or subscription of its own: each provider uses an app you install and sign in to yourself.

ProviderNeedsRunsModels
Tessera AIOllama + a modelOffline, on your GPUauto, or any installed tool-capable model
ClaudeClaude Code CLI, signed inYour Claude accountdefault, opus, sonnet, haiku
GPTCodex app (or codex CLI), signed in with ChatGPTYour ChatGPT accountdefault

Hover the provider menu to see what Tessera detected (for example “Claude Code CLI”, or a hint to install it). After installing a provider, restart Tessera so it is found.

Tessera AI (local, offline)

  1. Install Ollama for Windows.
  2. Pull at least one model in a terminal. Tessera prefers, in this order: qwen3-coder:30b, gpt-oss:20b, qwen3.6:27b, qwen3.5:9b, qwen3:8b, gemma4:e4b, qwen3.5:4b.
  3. Choose Tessera AI in the provider menu. Leave the model on auto, or pick one.
Recommended models
ollama pull qwen3-coder:30b   # best results, needs a GPU with lots of free VRAM
ollama pull qwen3.5:9b        # smaller: fits next to Unreal on most GPUs

You don’t need to start Ollama yourself: Tessera starts it hidden on first use. On auto, Tessera checks how much GPU memory is free right now and picks the first preferred model that fits (with some headroom); if none fits it falls back to the smallest preferred one.

Claude

  1. Install Claude Code (see claude.com/claude-code).
  2. Open a terminal, run claude once and sign in with your Claude account.
  3. Choose Claude in the provider menu, and optionally opus, sonnet or haiku.

Tessera runs the CLI headless with only the Tessera tools allowed. It looks for claude on your PATH and in %USERPROFILE%\.local\bin\claude.exe. Usage counts against your Claude plan.

GPT

  1. Install the Codex desktop app (or the codex CLI).
  2. Sign in with your ChatGPT account in the app.
  3. Choose GPT in the provider menu.

Tessera uses the CLI that ships inside the Codex app, read-only and limited to the Tessera tools. Usage counts against your ChatGPT plan.

Writing good prompts

  • Name every part you want. “A fairy cottage with a crooked chimney, a picket fence and two trees” works better than “a nice house”. The local AI plans one line per part.
  • Give sizes in metres when they matter: “a 40 m wide island”, “a 3-storey tower”.
  • Say where things go: “behind the house”, “to the left of the well”, “along the river”.
  • Ask for the output you need: “export a 16-bit heightmap for Unreal”, “send it to Unreal with live link”, “save it as harbour”.
  • Ask for controls: “expose the number of houses and the seed”. The AI names nodes so you can find the parameters that matter.
  • Iterate in the same chat: “make the roofs steeper”, “more trees on the north side”, “the chimney is floating, fix it”.
  • Mention node names if you know them: “use hf_erode”, “use fairy_village with 12 houses”. The node reference lists them all.

What happens when you send

  1. Plan. The local Tessera AI first writes a short numbered plan of every part, with sizes, positions and the node to use; it appears in the log. It is given only the nodes relevant to your request plus the closest proven recipe graph.
  2. Build. The AI adds, wires and sets nodes in batches with edit_graph. Each batch cooks the display node and reports errors, which the AI fixes.
  3. Look. The AI renders the result with Tessera’s built-in renderer (perspective, quad, top views, or a later frame for simulations) and judges it. Claude and GPT always can; the local AI does when its model can read images.
  4. Layout check. For assembled scenes, check_layout finds overlapping parts, floating parts, roofs not seated on their house and lifted chimneys, and returns the exact parameter fixes. The local AI repeats this until the scene is clean.
  5. Tidy. The graph is laid out, key nodes get comments, and a final null named OUT_… holds the display flag.

Stop, new chat, undo

  • Stop ends the current AI run. Nodes it already added stay in the graph.
  • New chat starts a fresh conversation. Without it, follow-up prompts continue the same conversation, so “make it taller” refers to what was just built.
  • Every AI edit is in the project’s undo history: Ctrl+Z steps back through it.
  • You can edit while the AI works, but it’s easiest to wait until it finishes, then tweak.

Teaching the local AI

The local model learns from recipes: proven graphs it can copy patterns from. The bundled examples are recipes, and you can add your own. Build or fix a project until it is right, then save it as a recipe through the API; it is stored in %USERPROFILE%\Tessera\recipes and used for future requests that resemble its title and description.

Save the active project as a recipe (PowerShell)
Invoke-RestMethod -Method Post -Uri http://127.0.0.1:7419/api/recipes/save -ContentType 'application/json' -Body '{"title":"Harbour town","description":"stone quay, boats, fairy houses along the water"}'