Getting started
Zephr is an MCP evidence memory. In local mode it runs entirely on your machine over stdio JSON-RPC, storing evidence in a single SQLite file scoped by worktree. Nothing leaves your machine in local mode: the server speaks stdio to the client that launched it, and the database is a file you own.
This section walks you through the flow end to end:
- Prerequisites and install — clone, install, build.
- Connect your client — point your MCP client at the server.
- Your first recall — capture and retrieve a belief.
- Claude Code memory plugin — plugin install, rollback, kill switch, and troubleshooting.
Prerequisites
Before you begin, confirm:
- Node.js 24.x — CI enforces Node 24;
engines.nodeis>=24 <25. - pnpm — the workspace installs with
pnpm install --frozen-lockfile. - An MCP client — Claude Code, Cursor, or any stdio JSON-RPC MCP consumer.
The only prerequisite that is easy to get wrong is Node — the workspace pins Node 24, and a mismatched runtime fails at build rather than at first tool call.
Important: admission is required
zephr-mcp is admission-scoped. Do not connect a bare zephr-mcp process with only
ZEPHR_DB_PATH for real work. Use the trusted launcher, which mints a single-use capability
and binding before any tool call is allowed. See Connect.