Skip to content
notdocs
notdocs
Use itFeaturesEverything notdocs does, from the editor and query blocks to merges, search and agents, with the page that explains each.AgentsGive Claude Code, Cursor or any MCP client the workspace: tokens, tools, resources and prompts.CLIEvery notdocs command: documents, import, comments, tokens and the organization, for a local folder or a hosted workspace.TemplatesEvery built-in template with the file it starts, from a task to a data processing agreement.PricingFree for one person, Solo and Team for a few. What each plan allows.
Move inSwitch from NotionImport a Notion export as typed markdown documents in your repository.Switch from ObsidianImport an Obsidian vault and keep working on the same files.notdocs vs NotionMarkdown in your git repository or pages in Notion's database: how agents reach each, and what each tool does better.notdocs vs ObsidianBoth keep markdown files. What differs is where they live, who edits them, and how.ChangelogWhat changed in notdocs, from the CHANGELOG in the repository.
notdocsAboutWhy the repository should be the workspace, and the decisions notdocs is built on.SecurityWhere data lives, how access is scoped, and how to report a vulnerability.ContactHow to reach notdocs for support, privacy, security and legal matters.
PoliciesLegalEvery policy and agreement that applies to notdocs.
Legal

AI disclosure

Where notdocs uses AI models, what they see, and how to turn them off.

Last revised

On this page

Where notdocs uses AI

Three features use language models, and every one of them is something you start.

  • The assistant in the workspace answers questions about a project and, when you ask, drafts or changes documents.
  • Inline rewrite in the editor rewrites the text you selected.
  • Search uses embeddings alongside full-text search, and the index is built whenever a document changes.

The web app is not itself an AI system; it is a workspace. There is no AI that reads your content on its own initiative.

You are talking to a model

When you use the assistant you are interacting with a language model, not a person. The panel says so, the model's name is shown in the picker, and answers arrive streamed with a cursor so it is clear they are being generated. Answers can be wrong, out of date, or confidently mistaken about your own documents; treat them as a draft.

What is sent to the model

Model calls go through Vercel AI Gateway to the provider of the model chosen. No account identifier travels with them.

  • Assistant. Your question, the conversation so far, and whatever the tools return: search hits, document lists, and the full text of documents the assistant reads or proposes to change.
  • /tasks. The full text of the open document.
  • Inline rewrite. The selection, up to 8,000 characters, and up to 4,000 characters of surrounding text.
  • Follow-up suggestions. The last question and answer.
  • Embeddings. Every document's title and body, in pieces of about 1,200 characters, whenever it changes.

Which models

  • Answers, rewrites and suggestions: deepseek/deepseek-v4.1-flash by default. You can pick any model the gateway offers in the panel; the choice is kept in your browser.
  • Embeddings: openai/text-embedding-3-small.

Model providers process the text under their own terms; Vercel AI Gateway publishes the retention and training terms that apply to each provider, and they differ by model, so check them for the one you choose. We do not train anything on your content ourselves.

Your own key

You can enter your own Vercel AI Gateway key in the assistant settings. It is stored in your browser, sent with your requests, and never stored on our servers. It applies to the assistant and inline rewrite, on every plan; on the Free plan it is the only way they run, and on a paid plan it takes over once the monthly allowance is used up. Calls made with your key are not metered by us. Embeddings always use the service's key.

Turning it off

  • The assistant and inline rewrite only run when you invoke them. Do not open the panel or the rewrite menu and nothing is sent.
  • Embeddings are built for every project of an organization whose plan includes semantic search (Solo and above); on the Free plan no text is sent to an embedding provider and search is full-text only. If a project on a paid plan must not have its text embedded, tell us at privacy@notdocs.dev and we will disable embeddings for it; full-text search keeps working.

What the assistant may change

The assistant can create documents, update them and move tasks between statuses. It asks before each of those and shows the change as a diff you can apply in part. What it writes is committed to your repository by the notdocs GitHub App with the commit message docs: change via assistant, so AI-made changes are identifiable in the history.

Labelling

Content the assistant writes into a document is not automatically marked as machine-generated inside the document, because you review and apply it deliberately. If you publish documents that contain assistant-written text where a reader could take it for a person's, label it; the EU AI Act asks that of deployers of generative systems.