Nick Khami (@skeptrune) launched Scribe on July 21st, a collaborative Markdown editor designed to let people and AI coding agents work inside the same live document.
Khami described Scribe as "like Google Docs, but for markdown" in a 32-post thread on X. The comparison covers the familiar parts: multiple cursors, real-time edits, comments, tracked suggestions and sharing. Scribe's narrower bet is that Markdown should remain the underlying file format and an agent should appear as an attributable collaborator instead of silently rewriting a file behind the editor. (scribe.skeptrune.com)
That distinction grew out of Khami's work with AI-generated plans, email copy and blog drafts. Moving those documents between coding tools and conventional collaborative editors made team review difficult, he wrote. Notion is more polished and supports features such as databases and exports, Khami told one respondent, while Scribe is built specifically for teams that want agents to operate directly on Markdown.
Khami is an engineering manager at Mintlify, which identifies Scribe on its site as a product "built by Mintlify Engineering." He joined Mintlify after it acquired Trieve, the search and retrieval infrastructure business he founded and led as CEO, in July 2025. Khami studied computer science and machine learning at the University of Texas at Austin and took Trieve through Y Combinator's Winter 2024 batch. (scribe.skeptrune.com)
Agents get identities and permissions
Scribe gives each agent its own API key and identity. A user can issue read, comment or edit permissions, and the agent's changes appear under its name in the document activity. Text that Scribe can attribute to an agent carries a persistent background tint, allowing reviewers to distinguish generated passages from human writing after the agent disconnects or the page reloads. (scribe.skeptrune.com)
The initial integration is distributed through a public agent skill repository. Installing it teaches Claude Code, Cursor and other compatible coding agents to create and search documents, edit content, and participate in comment or suggestion threads through Scribe's REST API. Users must generate a Scribe API key and expose it to the agent through an environment variable. (github.com)
The technical problem is concurrent writing. An agent cannot safely replace a document while several people are typing into it. Scribe requires version-aware API writes: an agent retrieves the current document version, then submits a patch against that version. The service merges the patch into the live editing session rather than overwriting intervening human changes. Comments are anchored to the selected text instead of fixed line and column positions, so they can remain attached as the surrounding document changes. (scribe.skeptrune.com)
Scribe's real-time layer uses Yjs through ProseMirror, with one Cloudflare Durable Object assigned to each document. The React client, REST API and collaboration layer ship through a single Cloudflare Worker, while D1 stores relational data and full-text search indexes and R2 stores uploaded images. Khami has exposed the architecture and local development instructions in Scribe's documentation, although the main application repository linked from those pages was not publicly accessible at launch. (scribe.skeptrune.com)
An AI-built product still needed human review
Khami said he built Scribe "100% with AI," then qualified the claim by pointing to the manual quality assurance and product judgment required to get it ready. The launch thread became a live bug queue: users identified sign-in friction, unclear human-versus-agent attribution, header behavior, checkbox handling and weak code-block contrast. Khami subsequently added a sandbox that works without signing in and posted fixes for several editor issues.
That sequence gives the launch claim its useful context. AI generated the implementation, according to Khami, while the product decisions, testing and correction loop remained human work. Scribe itself encodes the same division of labor: agents can write directly into the shared artifact, but their identities, permissions and changes stay visible to the people responsible for approving the document.
Scribe is available on the web, with desktop builds for Apple Silicon Macs and Linux. Its public pages did not list pricing at launch. Windows support was also absent from the published desktop documentation. (scribe.skeptrune.com)