Derive

Derive is a team library for AI-generated artifacts — HTML pages, Markdown documents, decks, and built sites — giving each piece of work a permanent URL, full version history, and a built-in review loop. Built by Derive Labs and licensed under the Functional Source License (FSL-1.1-ALv2), it is self-hostable as a single container and also available as a hosted service. The project reached its first tagged release, v0.2.0, in August 2026.

What It Is

Derive solves the problem of AI-generated work disappearing into chat threads, Slack pastes, and expired context windows. When an agent produces a page or document, Derive gives it a durable URL that resolves to every version ever published. Teams can comment on the exact text, request formal approvals, and @mention people or agents to act on feedback — all without the work ever leaving a single, searchable library. The product is described as "agent-native": it ships an MCP server so compatible clients like Claude Code, Codex, and Cursor can publish, read, and revise artifacts directly.

How the Publish-Review Loop Works

The core workflow has four steps the homepage labels Publish, Share, Collaborate, and Own:

• Publish: A CLI command (derive publish), HTTP API call, or MCP tool takes any HTML file, Markdown document, or multi-file build and creates a versioned artifact at a permanent URL. Every revision is a new version at the same link; no draft is ever discarded by default.
• Share: Artifacts can be scoped to the workspace, shared with anyone via link, or password-protected for external clients. Viewers never need an account; commenting and editing require sign-in.
• Collaborate: Comments pin to exact lines of text. @mentions work for both people and agents. A draft stays a draft until a named person formally approves a version, making approval status unambiguous.
• Own: Artifacts, versions, and comments export in open formats. The product runs as a single Docker container on SQLite and local disk by default, or on Postgres and S3/R2 at scale.

Agent and MCP Integration

Derive has an MCP server built in. One command connects compatible clients:

• claude mcp add --transport http derive https://derive.to/mcp for Claude Code
• npx skills add derive-to/derive --skill derive for any agent that reads skills
• A Cursor deep-link badge is provided in the README for one-click setup

The MCP server exposes tools for publishing, catching up on feedback, commenting, and reading any artifact or version. Agents are granted a role (viewer, commenter, or editor) and act within that scope. The homepage notes that agents do not consume billing seats — only human editors who publish work are counted.

Architecture and Self-Hosting

The repository is a TypeScript monorepo with a clear package split:

• apps/api — HTTP API, sandboxed artifact serving, viewer
• apps/web — web UI built with TanStack Start in SPA mode
• packages/core — domain logic: ports, publish pipeline, Markdown rendering, viewer shell
• packages/db — MetaStore supporting SQLite, Postgres, and Cloudflare D1
• packages/storage — BlobStore supporting local filesystem, S3, and R2
• packages/cli — derive init and derive publish commands
• packages/mcp — local stdio MCP server plus the agent skill guide

Every artifact is served on an opaque origin under a strict sandbox CSP, ships Open Graph and Twitter meta tags, and streams live updates over Server-Sent Events.

Update: Derive 0.2.0 — First Tagged Release

The GitHub repository shows the project was created in June 2026 and reached its first tagged release, v0.2.0, on August 12, 2026. The repository was last pushed on August 28, 2026, indicating active early development. The roadmap is itself a live Derive artifact at derive.to/roadmap, kept as a versioned document. The license is FSL-1.1-ALv2, which the project describes as converting automatically to Apache-2.0 two years after each release ships.