Turn your project's documents and spreadsheets into a versioned context build that AI agents read over MCP, with a citation on every answer.
Point Lorepack at a folder of runbooks, specs, spreadsheets and PDFs. It compiles them into an immutable build that Claude Code, Codex, VS Code or any MCP client can search and query, and every result names the file and lines it came from. Think Git for the context your AI depends on: see exactly what it reads, diff two versions, and roll back a bad update without rebuilding.
Status: pre-v0.1 and not yet on npm. Everything below runs today from source. Progress is tracked on the backlog.
Requires Node.js 24.15 or later.
git clone https://github.com/burakdede/lorepack.git && cd lorepack
corepack enable && pnpm install --frozen-lockfile && pnpm build
alias lore="node $PWD/packages/cli/dist/entry.js"After the v0.1 release it is npm install -g @lorepack/cli. See
Getting started for details.
lore dev ./my-docs # build the folder, serve it over MCP and HTTP, rebuild on change
lore connect claude-code # or codex, or vscode: wire up your agent and check it answers
- Answers with provenance. Every result names its file, heading path and lines. How provenance works.
- Spreadsheets stay tables. CSV and XLSX become typed tables an agent queries with read-only SQL, never flattened into prose. How tables are stored.
- Versioned context. Builds are immutable and content-addressed: diff any two, activate one, roll back without recompiling. The build lifecycle.
- Studio, a local inspector. See what was indexed and what was left out, what a model
would receive for a task, and copy any request as a CLI,
curlor MCP call. Take the tour. - Nothing else to install. No Python, Docker, model download, API key or account. Requirements.
- Deploy when ready. Project the same build to Cloudflare, verified before it goes live. Deployment.
- Getting started: install, the first build, and the full lifecycle
- Core concepts: why the build is the product, and how Lorepack differs from RAG servers and vector databases
- Connecting a client: Claude Code, Codex, VS Code, or any MCP client
- CLI reference
- Limitations: what v0.1 does not do
- All documentation
Start with CONTRIBUTING.md. Agents and contributors work to AGENTS.md, and vulnerabilities go through SECURITY.md.