Pensieve

A document centre your team keeps in git, and your AI plugs straight into. Push HTML documents to a repo; Pensieve serves them as a browsable workspace with full-text search and a link graph, and exposes that same workspace over MCP so Claude Code, Claude.ai, ChatGPT and anything else that speaks MCP can read it directly.

Git stays the source of truth, and sync is one-way — Pensieve never writes back. Documents change through the same pull requests as your code, and history is git log. Nothing lands in a proprietary store: if Pensieve went away tomorrow you would still have every document as a file you can grep.

Why HTML rather than markdown? Because documents deserve layout. Each page is a self-contained .html file — your styles, your components, no build step. AI writes it fast and browsers render it exactly. (.md is accepted too, and rendered on sync.)

How it works

• Workspaces — sign in with Google and you get a personal workspace. Create an organization, invite people, and everyone in it browses the same documents.
• Documents — plain .html files. <title> is the node label; optional <meta name="date"> and <meta name="tags"> drive sorting and search. CSS, JS and images sync alongside the documents, so relative references just work.
• Sync — install the Pensieve GitHub App on a repo and add sync sources (repo / branch / folder, each with its own mount prefix, several per workspace). Pushes sync via webhook; anything deleted upstream is pruned.
• Graph — links between documents become edges — root-relative and relative hrefs both resolve — and every rendered page gets its backlinks injected at the bottom. Nothing is stored: the graph is derived from links at read time.
• Search — ⌘K anywhere in the app. Full text over title, body and tags, Postgres-native: a weighted tsvector ranked with ts_rank_cd, plus pg_trgm for substrings and Chinese.

Connect your AI

Pensieve is a remote MCP server at /mcp.

claude mcp add --transport http pensieve https://pensieve.pathors.com/mcp

In Claude.ai, add the same URL as a custom connector. Either way you sign in with Google and approve a consent screen — OAuth 2.1, so the model sees exactly the workspaces you are a member of and nothing else.

Five tools: list_workspaces, search_documents, read_document, list_documents, related_documents. The last returns outgoing links and backlinks, so a model can walk the graph instead of only grepping it. Every result carries the document's canonical URL, so the answer cites a link a human can open.

Stack

Next.js (App Router) · better-auth (Google, organizations, MCP) · Drizzle + Postgres (Neon) · deployed to Cloudflare Workers via OpenNext. No queue, no cache layer, no vector DB — documents are rows.

Self-hosting

bun install
cp .dev.vars.example .dev.vars   # fill in DB + Google OAuth
bun run db:push                  # create tables
bun run dev

Deploy: bun run deploy (wrangler; set the same vars as secrets), or connect the repo to Cloudflare Workers Builds. Google OAuth redirect URI: <base-url>/api/auth/callback/google.

For GitHub sync, register a GitHub App (permissions: Contents read-only; webhook → <base-url>/api/github/webhook, push events), convert its key to PKCS#8 (openssl pkcs8 -topk8 -nocrypt -in app.pem) and set GITHUB_APP_ID / GITHUB_APP_PRIVATE_KEY / GITHUB_APP_WEBHOOK_SECRET.

Trust model

Documents are rendered as-is on the workspace origin: workspace members are trusted authors. Don't sync HTML you wouldn't run in your teammates' browsers. The MCP server is bound to the same membership — a connected model reads what the user who connected it can read, and nothing is writable through it.

MIT © Pathors