Configuration

The inf.yml file defines app settings and resource requirements.

Project Structure

1my-app/2├── inf.yml           # Configuration3├── inference.py      # App logic4├── requirements.txt  # Python packages (pip)5└── packages.txt      # System packages (apt) — optional

Basic structure

1namespace: myteam2name: my-app3title: My App4description: What my app does5category: image6kernel: python-3.1178resources:9  gpu:10    count: 111    vram: 24    # 24GB (auto-converted to bytes)12    type: any13  ram: 32       # 32GB

Fields

App tags

Tags describe what the app does and traits such as open-weights or realtime. They appear on the app record, in store listings, and in search indexing alongside name and description.

Declare them under version metadata in inf.yml:

1metadata:2  tags:3    - text-to-image4    - open-weights

belt app deploy --meta updates listing fields such as category and images; app tags are set on full deploy (or via the REST API), not through --meta alone.

In the web app, open your app → Info to edit category and tags. The tags field is a list: press Enter or type a comma to add a slug (up to 10). A later deploy that sends tags or declares metadata.tags replaces tags set only in the UI.

Store images

Optional URLs shown in the app Grid and store. The CLI reads these from inf.yml when you deploy:

1images:2  card: https://example.com/card.png3  thumbnail: https://example.com/thumb.png4  banner: https://example.com/banner.png5  dark:                       # optional: covers for the dark theme6    card: https://example.com/card-dark.png7    thumbnail: https://example.com/thumb-dark.png8    banner: https://example.com/banner-dark.png

Without dark, the three slots serve both themes. Omit images to use auto-generated covers after deploy.

To update the listing title, images, description, or category without shipping a new app version:

1belt app deploy --meta

Only non-empty fields in your local inf.yml are sent (for example, omit images.banner to leave the existing banner unchanged). The app must already be deployed at least once before --meta works.

App UI

Optional version metadata that names an artifact page to show beside finished app tasks, in the web app, in chat, and in MCP clients that support MCP Apps (see MCP server).

1metadata:2  ui:3    artifact: myteam/run-dashboard        # artifact ref namespace/name4    # artifact: myteam/run-dashboard@3    # pin one published version5    # artifact: art_abc123                # or artifact id

The artifact must exist and be readable by you at deploy time. The page must also be at least as visible as the app (a public app needs a public page), so everyone who can see the app can see its page. Otherwise deploy is refused with 400 validation_error naming the page. The API exposes the same object on AppVersionDTO.ui and in metadata.ui; see Apps API — custom UI page.

Publish the page first (belt artifact publish), then deploy the app with this block. See Artifacts — capabilities for what the HTML page may call (db, user, assets).

Resources

The CLI automatically converts human-friendly values to bytes:

• < 1000 → treated as GB (e.g., 80 = 80GB)
• 1000 to 1 billion → treated as MB (e.g., 80000 = 80GB)

1resources:2  gpu:3    count: 1        # Number of GPUs4    vram: 24        # 24GB5    type: any       # GPU type6  ram: 32           # 32GB

GPU Types

Note: Currently only NVIDIA CUDA GPUs are supported.

An app has one resources block. Setup parameters choose a model or configuration inside it, so the block has to fit the largest one. Models that need different amounts of memory are separate apps.

For CPU-only apps:

1resources:2  gpu:3    count: 04    type: none5  ram: 4

Categories

Dependencies

1torch>=2.02transformers3accelerate

System Packages (packages.txt)

For apt-installable system dependencies (both Python and Node.js):

1ffmpeg2libgl1-mesa-glx

Base Images

Apps run in containers with these base images:

Python environment

Each Python app version gets its own venv. It starts with build tools (setuptools, setuptools-scm, wheel, wheel-stub, build, packaging, ninja, scikit-build-core, cython, flit-core, numpy) plus pydantic and inferencesh.

torch and its CUDA wheels, accelerate, huggingface_hub, transformers and diffusers are not preinstalled. List every package your code imports in requirements.txt. transformers and diffusers need accelerate at runtime for device_map, low_cpu_mem_usage and enable_model_cpu_offload().

App versions that already have an environment keep it. New versions, and versions set up on a new worker, get this base.

Environment Variables

1env:2  MODEL_NAME: gpt-43  MAX_TOKENS: "2000"4  HF_HUB_ENABLE_HF_TRANSFER: "1"

Access in code:

1import os2model = os.environ["MODEL_NAME"]

Secrets and Credentials

Declare required secrets and OAuth credentials:

1credentials:2  huggingface:3    secrets: [HF_TOKEN]4    description: HuggingFace token for gated models56  google:7    capabilities: [google.sheets]8    description: Access to Google Sheets9    optional: true

Entries are keyed by the credential slug. The older list form (- provider: huggingface, - key: google.sheets) is still read. Use credentials: only; integrations: in inf.yml is rejected (renamed in CLI v1.18.59).

credentials: is the app-wide list. In a multi-function app, each function can declare the subset it uses; see Per-function credentials.

See Secrets and Credentials for details.

Next

→ Deploying