crate-cli npm package — Digcrate imports MCP tool servers from crate-cli rather than implementing them directly.
Each tool server exposes a set of MCP-compatible tools. The agent selects which tools to call based on the query, then executes them in sequence or parallel as needed.
Music databases
These are the core sources for music metadata, credits, samples, and discovery.
API keys required: Discogs (
EMBEDDED_DISCOGS_KEY + EMBEDDED_DISCOGS_SECRET) and Last.fm (EMBEDDED_LASTFM_KEY) need keys set in your environment. MusicBrainz, Genius, WhoSampled, and Bandcamp are accessed without authentication.
General knowledge
No API key required. Wikipedia is accessed via the public MediaWiki REST API.
Events
API key required: Set
EMBEDDED_TICKETMASTER_KEY in your environment. Get a key at developer.ticketmaster.com.
Connected (via Auth0 Token Vault)
These sources require the user to connect their account through Settings. The connection is managed by Auth0 Token Vault — Digcrate never stores raw OAuth tokens.Spotify integration requires Auth0 Token Vault to be configured (
AUTH0_DOMAIN, AUTH0_CLIENT_ID, AUTH0_CLIENT_SECRET). See Environment variables and Deployment for setup instructions.Images and media
No API keys required for these sources in standard use.
Web research
These sources extend the agent beyond structured databases into open-web research with citations.Exa, Tavily, and Perplexity each require their own API keys, which users configure in the Settings → API Keys panel within the app. These keys are stored encrypted in Convex and sent to the agent at runtime. They are not set as server-side environment variables.
Memory
Mem0 is available on the Pro plan. Users configure their Mem0 API key in Settings → API Keys.
Radio
No API key required. Radio Browser is accessed via the public community API.
How tool selection works
When a user sends a query, the agent receives the full list of available tools from all MCP servers. It decides which tools to call based on the query content and the system prompt. The agentic loop executes tool calls sequentially, feeding results back to the model until it has enough information to write a final answer. For research-heavy commands (e.g./influence, /story, /track), the agent gets a higher tool-call cap to allow deeper multi-source lookups. For conversational queries, it may use only one or two tools or none at all.
The tool execution log for each session is visible in the chat panel as collapsible “thinking” steps, showing which sources were queried and how long each call took.