What is DocGraph?

DocGraph is a local-first, HTML-native knowledge base with private AI: a desktop app where your notes, the links between them and an AI that has read all of them live in one place. It is built for a combination most tools make you choose between:

• An AI that can actually work. Not a chat box bolted on the side — an assistant that reads your whole vault, cites its sources, and — when you let it — can create notes, fill tables and draw whiteboards for you, showing you each change first.
• Nothing from your notes leaves your machine unless you choose it. The model runs locally by default. Network access is denied by default and shown to you live. There is no account and no telemetry. A cloud model (your own provider key), web search and the MCP server are all off until you turn them on, and DocGraph’s off-machine connections are logged. The one thing on by default is a daily update check, which sends nothing about you and can be turned off.
• Local RAG over the formats your work already lives in. Add your files and notes, ask questions, and get answers grounded in your own documents, with sources: PDFs, spreadsheets, Word files and notes, no importing required. With a local model nothing leaves your machine, so you can use AI on sensitive or confidential documents with complete privacy.

Those normally trade against each other. Capable AI usually means your documents go to somebody’s server. Real privacy usually means a tool that can’t do much. Broad format support usually means an importer and a database you can’t leave.

Underneath all three is one decision: your notes are plain files in a folder you own. A vault is just a directory on your disk — no proprietary database, nothing to export, nothing to escape.

AI-native, not AI-added

AI is off until you turn it on. Once it is on, it works across the whole vault:

• Ask the whole vault. Chat answers from your notes with a citation for each fact, and says so when your notes don’t cover it. Ask about one file, hand-picked notes, or everything.
• Find & open. “Open the note about my Berlin trip” finds the note and opens it — it understands “last time”, “next” and “only in Projects”.
• Agent Mode. It writes pages, fills tables, draws whiteboards, adds todos and reorganises folders. Every write is shown to you first; most can be undone.
• Writing help in the page. /write, /ask, /summarize, /continue, Beautify, selection rewrite, spell-check, and optional ghost-text completions — see AI in your documents.
• Search by meaning. Notes are embedded locally (English or multilingual) and results are re-ranked by a local model.
• It keeps up with the vault. Topic summaries answer big-picture questions, an optional check flags notes that contradict each other, and a memory file keeps standing facts across chats.
• Your choice of model. A built-in model, Ollama (including Apple MLX builds) or Apple’s on-device model — and OpenAI, Anthropic, OpenRouter or a custom endpoint only if you add a key. See Set up AI.
• Open to your other agents. An MCP server (off by default) lets Claude Code, Codex, Cursor and others read the vault and propose edits, which by default ask you before they write.

Why HTML first

Most tools treat HTML as an export format — something you get out at the end. DocGraph treats it as the source of truth: you edit the real file, and it is never flattened, normalized or converted.

What HTML buys you

Markdown is a lovely format for prose and stops there. HTML carries structure that a document actually needs once it stops being an essay:

• Typed tables with real columns, relations and computed values — not pipes and dashes.
• Live charts and saved queries that re-render from their data.
• Collapsible sections, so a long document reads as an outline you expand.
• Diagrams and embedded whiteboards, beside the prose that explains them.

That structure is not decoration. It is the difference between a document a machine has to guess at and one it can read precisely — which matters more now that the other reader of your notes is often a model.

This is a bet other people are making too

Engineers on Anthropic’s Claude Code team have publicly described moving their specs, plans and code reviews from Markdown to HTML — for the same reason: it is richer, still human-readable, and unambiguous for an AI agent consuming it.

• Why this Claude Code engineer uses HTML files as AI specs — Thariq Shihipar (Anthropic), video
• HTML is the new Markdown: how Anthropic engineers build with Claude Code — Lenny’s Newsletter
• According to Anthropic’s engineer, HTML is the new Markdown for AI agents — To Data & Beyond

DocGraph shipped on that conviction from day one. If you are writing specs for coding agents, this is the format they read best — and here your HTML stays the artifact itself: fenced code stays editable, Mermaid renders beside your prose, and an MCP server lets your own tools work the vault directly.

DocGraph is independent, and not affiliated with or endorsed by Anthropic.

And Markdown still works

None of this makes Markdown a second-class citizen. .md files edit inline the same way, are indexed and linked identically, and render Mermaid too. A few HTML-only things genuinely can’t exist in a Markdown file — see Markdown Support for exactly where the two differ and why.

Key ideas

• Read anything, edit what matters. PDFs, spreadsheets and Word documents open, index and answer questions alongside your notes, with sources and no importing required — read-only, because the editor is for HTML and Markdown.
• Linked thinking. Connect notes with links and #tags, and explore the connections in the link graph.
• Whiteboards. Infinite canvases for visual thinking — sticky notes, shapes, pen drawing, and ERD table nodes for database design.
• Local AI, off until you want it. Chat with your vault using a model that runs on your computer (Apple’s on-device model, DocGraph’s built-in runtime, or Ollama if you already use it), and let it act on your notes with your confirmation. A cloud provider is an option only with your own key. Off by default — DocGraph won’t download a model or index anything for AI until you say so.
• Autosave everywhere. There is no save button. Edits are written to disk automatically as you work.
• Version history, when you want it. Turn on versioning and DocGraph keeps a git-backed history of the vault that you can browse and restore from — see Version History.
• Plugins, coming soon. Plugins that extend DocGraph are planned for a later release. See Plugins.
• Yours to inspect. Every network call, every AI run and every secret found is recorded in the app — see Privacy & Security for what’s covered and what isn’t.

Free to use

DocGraph is free.

Ready to try it? Head to Installation.

Get in touch

Questions, bug reports, or something in these docs that doesn’t match the app? Write to contact@docgraph.app.