Person OS for your static exports.

A local-first personal & org knowledge-graph library. Ingests Gmail mboxes, Drive folders, and other archive exports into a typed bitemporal graph. Agents navigate it over MCP.

View on GitHub → Try it locally

v0.1.0-alpha· Apache 2.0· 27 tests passing· TypeScript / Bun

What it does

Three primitives, one unifying bet: the archive is the source of truth because logins rot.

01 · Ingest

Static-exports-first

Gmail mbox archives, Google Drive folders, and (soon) Slack, Notion, and Microsoft 365 exports.

Not live APIs. The archive outlives the login.

02 · Model

Typed temporal graph

Two primitives — Node and Edge with open kind strings.

Bitemporal from day one (tx_created, tx_invalidated, valid_from, valid_to) with provenance on every record.

03 · Surface

MCP-native

Seven intent-shaped tools your agent can call from Claude Desktop, Cursor, or any MCP client.

Each return carries inline provenance — never “the LLM made it up.”

The MCP surface

A small, intent-shaped tool set. Six tools for navigating, one for writing — all returning structured content with cited sources.

searchFTS5 + filter over the graph (text, kinds, min_confidence).

neighborhood1-3 hop subgraph around a given node.

timelineTime-ordered events for a subject or window.

whyWalk the provenance tree back to ingested leaves.

fetchResolve hypha://node/{id} or hypha://edge/{id} URIs.

recordAgents write — subject to Cedar policy + idempotency keys.

askNL question → compiled StoreQuery (Claude Haiku + FTS fallback).

Try it locally

Clone, install, ingest your own archive, and serve it to Claude Desktop. No accounts, no API keys (except optional ANTHROPIC_API_KEY for NL queries).

# Clone and install
$ git clone https://github.com/nalediym/hypha && cd hypha
$ bun install

# Ingest a Gmail mbox archive
$ bun run hypha ingest gmail-mbox path/to/archive.mbox

# Run inferrers — identity resolution, DLP scanning
$ bun run hypha infer identity-resolver
$ bun run hypha infer dlp-scanner

# Search the graph
$ bun run hypha search "Dinner plans"

# Serve over MCP for Claude Desktop / Cursor
$ bun run hypha serve

Privacy by default

Hypha runs on your machine. Nothing about your archive leaves it unless you explicitly ask.

Stays on your machine. Ingested data, the SQLite store, identity-resolver clusters, and DLP findings — all written to .hypha/ in your project directory.
No accounts. No telemetry. No analytics SDK, no remote logging, no “phone home” check-ins.
Local-first by default. SQLite + FTS5 + sqlite-vec for embeddings. Cedar policy interface for governance.
Optional outbound: the ask tool calls Claude Haiku for NL → StoreQuery compilation when ANTHROPIC_API_KEY is set. Omit the key and it falls back to FTS-only.

Status

Shipped v0.1.0-alpha on April 22, 2026. Core wired end-to-end; some adapters and the web UI are deferred to future milestones — see the CHANGELOG for the honest scope.

Adapters

gmail-mbox, gdrive-folder

Inferrers

identity-resolver, dlp-scanner

Surfaces

stdio MCP, read-only HTTP

Tests

27 passing, 1 skipped