Modules
Embedded offline context store for AI agents (OpenViking-inspired).
The on-the-wire shape of a transfer archive: its members, its manifest, and the validation every entry must pass.
A bounded tar archive, optionally gzipped, in both directions.
The local model capabilities the store depends on.
Durable state the context store keeps, expressed without reference to any particular database.
An outer surface that exposes the store's public facade to a protocol.
The durable job queue: the SQL behind AgentDb.Core.Storage's work
operations.
Loads models through Bumblebee and reports which serving runs them.
Manages embedding and LLM models for vector search and summarization.
Runtime that loads embedding and summarization models and runs inference.
Apple MLX backend (EMLX + EMLXAxon).
Bumblebee + EXLA backend (CPU by default, CUDA/ROCm when configured).
State struct for ModelManager
The MCP tool inventory over the store facade.
Stateless runtime diagnostics: low-cardinality telemetry, distributed traces, and redacted structured logs.
Holds the store's recent failures and outcome counts, so an operator can see what has been going wrong without tailing a log.
Deterministic demo dataset for development and test environments.
Reading an Agent Skills bundle, and refusing one that cannot be stored as it is.
The main web module for AgentDb.
The shared behavior of a console page.
The console's documents page: what the store holds, one page at a time, and how to find a document in it.
The console's memories page: what the store holds as memory.
The console's overview: what the store is doing right now.
The console's sessions page: looking a session up by id.
The console's skills page: importing Agent Skills into a user's subtree, plus the installed-skill inventory.
The console's storage page: what the store holds and how much of it is cache rather than content.
The console's sections, as function components.
Optional bearer-token authentication for the API.
CORS for the API.
The WebSocket surface.
The transport's view of the store.
Documents over HTTP.
Liveness, as the store reports it.
MCP Streamable HTTP endpoint.
What the store's models are doing, including a load in progress.
Search over HTTP.
Sessions over HTTP.
Editing one document.
The HTTP and WebSocket listener.
Compatibility shim over AgentDb.Mcp.
The HTTP surface.
Authentication for the browser surface.
The WebSocket endpoint.
Mix Tasks
Reads a document's L0 abstract -- the one-line answer to what this document is about, for deciding whether to read the content at all.
Checks store health and configuration.
Exports store data to a tar archive.
Discovers files and directories whose URI path contains TERM.
Searches full document content (L2) for a literal substring.
Imports store data from a tar archive.
Imports Agent Skills from the command line.
Indexes Elixir sources structurally.
Reads a document's L1 overview -- the bounded scan of its key points, for deciding whether to read the content at all.
Reads a document's full content (L2).
Reads memories back, optionally scoped.
Searches the context tree.
Seeds a deterministic demo dataset for dev and test environments.
Shows a depth-limited projection of the context tree.