Getting started
Zephr is a local-first MCP server. It runs on your machine via stdio JSON-RPC, stores evidence in a SQLite file per worktree, and exposes exactly eight tools to your coding agent. No account, no cloud, no telemetry.
Prerequisites
- Node.js 22.x
- A supported MCP client (Claude Code, Cursor, or any stdio MCP consumer)
Install
Zephr ships as a single npm package with native sqlite-vec bundled per platform:
npx zephr-mcpOr install globally for persistent access:
npm install -g zephr-mcpConfigure Claude Code
Add Zephr to your Claude Code MCP configuration. The server runs over stdio and needs a database path (defaults to :memory: if unset):
{
"mcpServers": {
"zephr": {
"command": "npx",
"args": ["zephr-mcp"],
"env": {
"ZEPHR_DB_PATH": "~/.zephr/store.db"
}
}
}
}Configure Cursor
Cursor supports MCP servers via its settings. Add the same configuration to your Cursor MCP config file:
{
"mcpServers": {
"zephr": {
"command": "npx",
"args": ["zephr-mcp"],
"env": {
"ZEPHR_DB_PATH": "~/.zephr/store.db"
}
}
}
}ZEPHR_DB_PATH env var sets the primary database location; Zephr creates per-worktree tables within it.The eight tools
Zephr exposes exactly eight MCP tools. This surface is frozen — a ninth tool is never added (ADR-Z-06). New capability routes through action enums or optional arguments on existing tools.
zephr_remember— capture a belief with provenancezephr_recall— hybrid lexical + semantic retrieval with abstentionzephr_why— explain provenance, versions, contradictionszephr_verify— T1 structural and T2 consistency checkszephr_session— session lifecycle and handoff summarisationzephr_rules— read/write review and contradiction policieszephr_status— service identity, capabilities, storage statszephr_admin— capability-gated stats, purge, migrate
Admin token (optional)
Destructive operations (zephr_admin.purge, zephr_admin.migrate) require a capability minted from ZEPHR_ADMIN_TOKEN:
export ZEPHR_ADMIN_TOKEN="your-secret-token"
npx zephr-mcpWithout the token, zephr_admin.stats (read-only) still works. Purge and migrate are refused.
Architecture
Zephr uses a local SQLite database with FTS5 for lexical search and sqlite-vec for semantic retrieval. The hybrid fusion uses reciprocal rank (RRF). Every recalled belief carries a full provenance chain and bi-temporal recording. The store, read view, and admin capability are separated per ADR-Z-02.