Open-source design agent: on-brand stills, decks, diagrams, charts and motion graphics from a chat thread, built on Deep Agents.
Models
Design Agent talks to models through LangChain, so any chat model with tool calling and vision can
run it. Set one in .env (or uv run design-agent config set …):
DESIGN_AGENT_MODEL=anthropic:claude-opus-5-5 # provider:model
ANTHROPIC_API_KEY=…
provider:model is LangChain's init_chat_model form. Anthropic and OpenAI are installed by
default; Google, Groq and Mistral are extras (uv sync --extra google). The LangSmith Gateway is an
optional preset (langsmith:anthropic/claude-opus-5-5 with LANGSMITH_GATEWAY_API_KEY, else
LANGSMITH_API_KEY), never required. Anthropic and OpenAI models go through the gateway's native
routes and keep every feature in the table below; other providers use its shared route. Images
come from the first of these that is set: OPENAI_API_KEY (OpenAI), GOOGLE_API_KEY (Google's
Gemini image models), or, on a deployment that uses the gateway (a langsmith: model or
LANGSMITH_GATEWAY_API_KEY), the gateway: its Google path when the lead model is Google, its OpenAI
path otherwise, with the provider key stored in that gateway workspace.
DESIGN_AGENT_IMAGE_MODEL=gemini-… (for example gemini-nano-banana-2.1) picks Google whenever
Google is reachable. Google's image models make no transparent backgrounds. For an OpenAI-compatible endpoint, set DESIGN_AGENT_MODEL_BASE_URL and
DESIGN_AGENT_MODEL_API_KEY_ENV (the name of the variable that holds its key).
Roles
| Role | Setting | Default | What it does |
|---|---|---|---|
| Lead | DESIGN_AGENT_MODEL, DESIGN_AGENT_MODEL_EFFORT |
anthropic:claude-opus-5-5 at high |
the conversation, plans, creative calls |
| Helpers | DESIGN_AGENT_HELPER_MODEL, DESIGN_AGENT_HELPER_EFFORT_CAP |
the lead's model; effort per task (low, xhigh, max) |
background builders, critics, renders |
look |
DESIGN_AGENT_LOOK_MODEL |
the lead's model at low |
looks at renders (up to 8 images per call) |
| Tool selector | DESIGN_AGENT_TOOL_SELECTOR_MODEL |
the lead's model | picks MCP tools, only for models without native tool search |
| Learning | DESIGN_AGENT_LEARNING_MODEL |
a low-effort helper | the post-job review that proposes observations |
Every role needs vision. uv run design-agent doctor checks each one against the registry.
The capability registry
src/design_agent/models/registry.py records, per exact model id, what the agent may rely on:
vision, effort levels, context and output limits, prompt-cache lifetimes, server-side context
editing, and how tools are disclosed. Only verified entries claim provider features; anything else
runs on the portable path (all features that do not depend on the provider).
| Model | Tool disclosure | Notes |
|---|---|---|
anthropic:claude-opus-5-5, claude-fable-5-1, claude-sonnet-5-5 |
provider tool search | efforts low…max; adaptive thinking; 5 m / 1 h cache |
anthropic:claude-haiku-5-5 |
provider tool search | no effort parameter; good for look and the selector |
openai:gpt-6.1-sol, gpt-6-sol, gpt-5.6-sol |
provider tool search | Responses API; reasoning effort |
google_genai:gemini-3.6-flash |
bounded selector | no provider-side tool search |
| anything else, the gateway's shared route, or a base URL | bounded selector | no provider-feature claims |
Adding a model: add an entry with the facts from the provider's current documentation, a test in
tests/unit/test_config_and_registry.py, and run the eval set (docs/evals.md) before relying on it.
Cost and limits
- Each call is bounded by
DESIGN_AGENT_MODEL_TIMEOUT_SECONDS(600 by default); provider SDK retries are off and the agent's own retry handles transient failures (failure boundaries). - The lead's prompt cache lives an hour (
DESIGN_AGENT_CACHE_TTL), helpers' five minutes. - Spend is estimated per job from
config/prices.toml;DESIGN_AGENT_JOB_BUDGET_USDmakes the agent ask before going past a budget.