What Is NeverWrite? A Private, Local-First Knowledge Workspace for Small Teams

NeverWrite is a local notes and documents workspace for small teams, with AI assistance that proposes changes but does not apply them until you approve. It's free, open source, and requires no account, so it suits teams that want their knowledge base stored as local files rather than in a vendor's cloud. It is not a fit if you need real-time multi-user editing or a hosted service you never have to maintain.

What NeverWrite actually is

The site describes it as a "private knowledge workspace for small teams" built on "local notes and documents." Two things follow from that:

• Your notes and documents live as files you control, not as records inside a proprietary database.
• The AI layer sits on top of those files and works with them, rather than replacing them.

The tagline — "Your knowledge base, made smarter" — frames it as an improvement to a knowledge base you already keep, not a new place you have to move everything into.

The core mechanism: AI that asks first

The distinguishing feature is the approval gate. When the AI works on your documents, it produces edits as pending changes. Nothing is written to your vault until you accept them.

The product page shows this in practice. Asked to review a file called q3-launch-plan.md, the assistant reads two files, then reports:

"I clarified the approval gate in q3-launch-plan.md and drafted launch-checklist.md for the final handoff. The customer-facing claim is unchanged. Both edits are pending — nothing is written to the vault until you keep them."

The interface then lists the edits — q3-launch-plan.md +8 -3 and a new file launch-checklist.md +12 — with a diff view and an "Accept Edits" control. The instruction given in that session was explicit: "Apply both as a review. Keep every edit pending until Maya approves it."

That pattern is the point. You can let an agent read and propose freely while keeping a human decision between the proposal and the file on disk.

What the AI can do in a session

From the same example, the assistant:

• Read multiple files before acting (the session notes "Read 2 files").
• Edited an existing document and created a new one.
• Explained its reasoning, including what it deliberately left unchanged ("The customer-facing claim is unchanged").
• Held all changes for review rather than committing them.

The interface also shows a model selector (the example uses "Sonnet 5.5" at a "High" setting), a message box with @ to include context and / for commands, and a history of past agent sessions with timestamps.

How teams use it

The example workspace is organized around clients and projects: a sidebar shows Clients / Fieldwise / Launch, with documents like fieldwise-audience-research and fieldwise-approval-flow alongside the launch plan. Documents carry properties and status fields ("Set status"), and there's a search box over the workspace.

A realistic small-team workflow looks like this:

1. Keep project and client documents as local files in a vault.
2. Point an agent at the relevant files — for example, "Review q3-launch-plan.md. Keep the scope firm, make the approval path explicit, and don't change the client promise without support from client-discovery.md."
3. Read the proposed edits and the reasoning behind them.
4. Accept, reject, or adjust before anything is written.

The constraint you set in step 2 matters. In the example, the instruction not to change the client promise without support from another file is what kept the customer-facing claim intact — the AI reported that explicitly.

Who it fits, and who it doesn't

Two caveats worth stating plainly. First, "free and open source, no account required" is what the site says; the material contains no pricing page, payment links, or paid tiers, so treat free as the current described state rather than a permanent commitment. Second, local-first means you own backup, sync, and access control — that's the trade for privacy and file ownership.

The practical test

If your team already keeps notes and documents as files, and the thing stopping you from using AI on them is that you don't trust it to change things behind your back, NeverWrite is aimed directly at that problem. The approval gate is the feature to evaluate first: open a real document, ask for a change, and check whether the pending-edit review gives you enough context — the diff, the reasoning, and what was left untouched — to make a confident decision.