Documentation

How Backpack fits together.

Local tools, persistent graphs and a viewer for understanding the result.

The components

  • backpack-ontology: the graph engine and local MCP server. Your AI client calls its tools to read and write knowledge.
  • backpack-viewer: a local web application for browsing and editing your graphs.
  • backpack-ontology-plugin: optional Claude Code guidance for building, researching and refining graphs.

Follow the local quickstart →

Where your data lives

By default, graph data lives under ~/.local/share/backpack and configuration under ~/.config/backpack. The tools also respect the XDG data and configuration directory settings.

Set BACKPACK_DIR to use a different root: the engine stores configuration in BACKPACK_DIR/config and data in BACKPACK_DIR/data. Give the MCP server and viewer the same setting.

Graphs use an append-only event log with materialized snapshots, branches and version history. Back up the whole Backpack data directory rather than only a snapshot. Stop writers while taking a simple file copy so the backup is consistent.

A useful first workflow

  1. Create a graph for a bounded topic or project.
  2. Add entities, relationships and the evidence behind them.
  3. Ask your client to read back the graph and check the result.
  4. Explore it in the viewer, then refine it as you learn more.

Graph entries are only as reliable as their sources and the decisions of the person or model creating them. Keep sources and uncertainty alongside important claims.

Local storage and privacy

Local storage does not mean your AI client is offline: a hosted model may receive the graph content returned by MCP tools. Review your client's data policy and any extensions you enable.

The published engine includes optional telemetry. Our quickstart disables it with DO_NOT_TRACK=1. See local software privacy.

About the former hosted service

Hosted accounts, billing, remote MCP, cloud sync and hosted sharing have been retired. The website does not provide those services. Older releases and documentation may still mention them; use the local setup above.