DEVON - Discovery Engine and Vault for Open Neural models

"DEVON manages the models. KITT tests them."

CLI tool, REST API, and Web UI for discovering, downloading, and managing LLM models from HuggingFace and other sources.

Full Documentation | CLI Reference

• Smart search — filter by provider, size, parameters, format, task, license
• Easy download — by URL or model ID with automatic resume
• Local vault — organized storage with portable manifest and disk usage tracking
• Directory scanning — discover models added outside Devon (custom fine-tunes, manual copies) with automatic metadata inference
• KITT integration — export model paths for inference testing
• Source plugins — extensible architecture for model sources
• YAML configuration — deep-merged config with sensible defaults
• Web UI — browser-based dashboard for search, downloads, model management, and configuration
• REST API — FastAPI server for remote model management
• Docker ready — containerize with a single volume mount (Web UI included)

# Install
poetry install
eval $(poetry env activate)

# Search for models
devon search --provider qwen --params 30b --format gguf

# Download by URL
devon download https://huggingface.co/Qwen/Qwen2.5-32B-Instruct

# List downloaded models
devon list

# Discover models added outside Devon
devon scan

# Export for KITT
devon export --format kitt -o models.txt

The search command supports these filters (combine freely, AND logic):

devon search "query"                          # text search
devon search --provider qwen                  # by author/org (-p)
devon search --params 30b                     # by parameter count (±20% tolerance)
devon search --size "<100gb"                  # by file size (<, <=, >, >=)
devon search --format gguf                    # by format (-f: gguf, safetensors, pytorch, onnx)
devon search --task text-generation           # by pipeline tag (-t)
devon search --license apache-2.0             # by license (-l)
devon search --limit 50                       # max results (default 20)

Filters also work inline: devon search "qwen 30b gguf" auto-extracts params and format.

See the full filter guide for detailed examples and sample output.

Config file at ~/.config/devon/config.yaml. Only override what you need:

storage:
  base_path: /mnt/data/models
  max_size_gb: 500

See the full configuration guide for all options.

Install the API extras, then start the server:

poetry install --extras api
devon serve                       # http://127.0.0.1:8000
devon serve --host 0.0.0.0 --port 9000

DEVON uses three-tier authentication on /api/v1/* endpoints (checked in order):

On first launch with no API key configured, the Web UI walks you through setup at POST /api/v1/setup, which generates and persists a key automatically.

# Production — require a bearer token
DEVON_API_KEY=mysecretkey devon serve
curl -H "Authorization: Bearer mysecretkey" http://localhost:8000/api/v1/models

# Local development — explicitly disable auth
DEVON_API_KEY=disable devon serve
curl http://localhost:8000/api/v1/models

Security note: Environment variables are visible in /proc/<pid>/environ and in ps e output. On shared hosts, prefer injecting DEVON_API_KEY via a secrets manager or a file-based mechanism rather than a shell export.

# Health check
curl http://localhost:8000/health

# Search for models
curl "http://localhost:8000/api/v1/search?provider=qwen&limit=3"

# List local models
curl http://localhost:8000/api/v1/models

# Start a download (returns 202 Accepted with job ID)
curl -X POST http://localhost:8000/api/v1/downloads \
  -H "Content-Type: application/json" \
  -d '{"model_id": "Qwen/Qwen2.5-1.5B"}'

# Check download progress
curl http://localhost:8000/api/v1/downloads

# Storage status
curl http://localhost:8000/api/v1/status

Devon includes a browser-based dashboard for managing models without the CLI. When devon serve starts, the Web UI is available at the root URL (e.g. http://localhost:8000/).

Pages:

On first launch (no API key configured), the Web UI presents a setup flow that generates and stores an API key automatically. Devon works immediately with defaults — configuration is recommended but not required.

The Web UI is built from frontend/ (React + Vite + TypeScript + Tailwind CSS). The Docker image builds it automatically. For local development:

# Build the UI into src/devon/ui/static/
make build-ui

# Or run the Vite dev server with API proxy
make dev-ui    # UI at http://localhost:5173, proxies /api to :8000

docker compose up -d
curl http://localhost:8000/health   # API health check
# Open http://localhost:8000/ in a browser for the Web UI

Mount your existing models directory:

DEVON_DATA_PATH=/mnt/models docker compose up -d

The container stores models (and the manifest) at /data/models/ and config at /data/config.yaml. A single -v /your/path:/data covers everything.

Note: The default configuration runs a single uvicorn worker to avoid race conditions on the JSON index file.

• KITT — LLM inference testing suite (docs)

Apache 2.0