Zephr documentation
Zephr is a local-first, trustworthy memory and evidence graph for coding agents and AI-assisted development. It gives an agent a durable store of what it has learned — "beliefs" — each one anchored to provenance (the source, anchor, and content hash it came from) and carried through an immutable, verifiable version chain.
The MCP server (@zephr-ai/mcp, binary zephr-mcp) exposes Zephr's tools to any MCP-compatible
client over stdio, scoped to an admitted worktree established by a trusted launcher.
Where to start
- New to Zephr — read the Getting started section: prerequisites, install, connect your client, and capture your first belief.
- Understand the model — Concepts covers architecture, beliefs, provenance, and trust.
- Configure a client — How to configure your client shows the exact MCP server configuration shape.
- API surface — the Reference section lists the MCP tools and status commands.
The trust model at a glance
- Admission-scoped — tool calls that pass a
worktreeIdoutside the admitted scope are rejected withscope_violation. - Trust Firewall (
zephr_rules) — per-worktree policy on review state, contradiction handling, content bounds, and recall eligibility. It can only ever narrow retrieval. - Version-chain integrity — current-version pointer, contiguous revisions, and non-overlapping transaction intervals are verified, not assumed.
- Contradiction & supersession — beliefs that contradict or supersede one another are tracked with directionality.
Distribution status
Zephr is not yet published to npm. The working install path is build-from-source from the repository. Linux x64 is the only CI-proven platform today; Windows and macOS are not claimed supported yet. See Install and Self-hosting.