Configuration
Everything here is optional: the compose file from Getting started already turns on every Tool.
- API keys go in the app's Settings panel, or in a file named
.envnext todocker-compose.yml. - Other options go in
docker-compose.ymlitself, under theagentservice'senvironment:.
Run docker compose up -d after changing either file.
API keys
| Key for | In Settings | In .env |
|---|---|---|
| Anthropic (Claude) | Chat model and API keys | ANTHROPIC_API_KEY |
| OpenAI | Chat model and API keys | OPENAI_API_KEY |
| Google (Gemini) | Chat model and API keys | GEMINI_API_KEY |
| Materials Project search | Databases | MP_API_KEY |
A key in .env takes precedence over one saved in Settings. Settings saves
keys inside the container, in ~/.config/goldilocks/config.toml, which is
kept in the goldilocks_home volume. A key is only ever sent to the service
it belongs to.
A Materials Project key is free from materialsproject.org/api. Without one, Find in Databases searches the other databases and skips Materials Project.
The model
Pick the model in the menu next to the message box. Cloud models need their
API key; the local model, Qwen3.8-27B, needs Ollama
running on your computer with ollama pull qwen3.8 done (see
Getting started).
To use one model for every chat whatever the menu says, add it under
environment: in docker-compose.yml as a
litellm model name:
GOLDILOCKS_AGENT_MODEL: anthropic/claude-sonnet-5
Ollama is expected on your computer at its default port, 11434. The compose
file reaches it as http://host.docker.internal:11434 (OLLAMA_API_BASE).
Options set in the compose file
docker-compose.yml sets these for you:
| Option | Set to | What it does |
|---|---|---|
GOLDILOCKS_AGENT_CORE_AUTOSTART |
1 |
Turns on DFT Workbench: Goldilocks installs goldilocks-core from PyPI the first time it is needed and starts it. |
UV_TORCH_BACKEND |
cpu |
Installs goldilocks-core's PyTorch for the CPU, which is smaller and faster there than PyPI's default GPU build. |
GOLDILOCKS_AGENT_MLIP_ENABLED |
1 |
Turns on MLIP Playground's calculations. The first one installs its MACE environment, a few GB. |
The DFT parameters themselves always come from goldilocks-core, which has its own documentation.
Running the website
The shared website at workbench.goldilocks.ac.uk
uses the goldilocks-agent-web image and options of its own (anonymous
visitors, chat retention, a pinned model). They are described in
deploy/stfc-cloud/README.md.