Most collaboration tools pick two of three: multiplayer editing, AI agent integration, or local plain files. Quire does all three. The open-source tool, now in public beta, turns a folder of Markdown files into a live multiplayer workspace where humans and AI agents edit the same documents in real time, with every change attributed, visible, and separately revertable. The filesystem stays authoritative throughout.
There is no import step, no cloud copy, no conversion round trip. Point Quire at a folder and the .md files become the workspace. People edit in vim, VS Code, or the browser. AI agents join the same CRDT session through an MCP endpoint. Git pulls and external tool edits merge as deltas, not overwrites. When you stop Quire, the Markdown files and a .quire/state/ directory are all that remain. Delete the state directory and you lose only the collaboration layer.
Why the agent model matters
The difference between Quire and every other Markdown collaboration tool is how agents participate. In most setups, an AI agent reads a file, makes changes in isolation, and writes a diff back. The human reads the result and decides what to keep. The agent is a batch process, not a collaborator.
In Quire, the agent is a peer in the same CRDT session. It has a visible cursor, attributed spans, and a suggestion mode where edits appear inline but never reach disk until accepted. A person can type directly through the agent's edits without conflict. The agent gets a presence identity, and every character it places carries its author tag. Revert by author removes one agent's contributions without touching adjacent human-written text.
This is not a theoretical feature. The MCP endpoint connects directly to Claude Code, Codex, Cursor, or any MCP-speaking client. The agent receives tools to list documents, read content, edit with suggest mode, append, manage suggestions and comments, and search the vault. The connection is local by default, bound to 127.0.0.1, with no account or credential required on the Quire side.
Plain files as source of truth
Cloud editors that import and export Markdown treat the cloud document as the workspace. The local file is a snapshot you reconcile after the fact. Quire inverts that model. The files on disk are authoritative. Changes from editors, agents, and git merge into the live session continuously.
The release suite validates byte-for-byte stability of syntax-rich Markdown through editing sessions. The test fixture covers YAML frontmatter, task lists, tables, fenced code, Mermaid diagrams, wiki-links, footnotes, raw HTML, and comments. A separate reproducible comparison shows what survives a Google Docs import and export round trip. The result: the same fixture loses fidelity through the conversion, while Quire preserves it without touching the source.
For teams that version-control their documentation, this matters. Git pulls during an active session merge as deltas. The filesystem watcher picks up files created by agents, external tools, or Discover without requiring a refresh. The Markdown stays clean and parseable by every tool in the existing workflow.
What runs on your machine
Quire starts with a single command:
npx quiredocs ~/my-notesThe server binds to 127.0.0.1 and opens a browser at http://127.0.0.1:4321. Core editing makes no outbound requests. The npm package has no runtime dependency downloads and no preinstall, install, or postinstall hooks. There is no analytics, no advertising SDK, no auto-updater, no credential prompt, no background service, and no payment integration.
The security model is explicit about what runs and what does not. Files reads and writes are scoped to the folder passed on the command line. Git snapshots are off by default and only commit paths Quire changed. Discover contacts GitHub APIs only after a person searches, previews, or installs content. Code execution is off by default and requires a person to explicitly run a fenced block. Sharing is local only, served by the running process. The server rejects cross-origin browser requests.
The tool has no authentication. Treat the bound address as the access boundary. Running on 0.0.0.0 means anyone who can reach the port can read and edit every document in the vault. The threat model and known limitations are documented in SECURITY.md.
Attribution and provenance
Every span of text in a Quire document carries its author. Toggling the Authors view tints text by who wrote it. Comments anchor to text ranges and survive concurrent edits, flagged rather than dropped when their anchor is deleted. Authorship, comments, provenance, and policy persist across restarts in the .quire/state/ directory.
The provenance receipt feature turns a document into a shareable HTML page. It shows how much a person wrote, how much an agent contributed, which suggestions were accepted, and includes a replay of the document being written. The HTML is self-contained, works offline, and requires no server or account. For teams that need to audit AI-generated content or demonstrate human oversight, this provides a concrete artifact.
Review requests work through capability links scoped to a file or vault, with view, comment, or edit permissions enforced server-side. The reviewer comments without an account. The server restricts comment-only mode at the protocol level, not by hiding the editor interface.
Competitive position
The landscape splits into tools that handle two of the three requirements. Notion, Outline, Docmost, and HedgeDoc offer multiplayer collaboration and AI integration but store documents in their cloud. Obsidian, SilverBullet, and Logseq work with plain files and have some agent support but lack real-time multiplayer. SoloMD supports agents and plain files but no multiplayer. CollabMD offers multiplayer and plain files but no agent integration.
Quire is the only tool that does all three. The agent is not a batch process whose diff you read later. It is a peer in the same editing session with presence, attribution, policies, suggestions, and author-specific revert. This is the model that matters when humans and AI agents share authorship of the same document.
The Discover feature extends the workflow beyond a single vault. It indexes widely used Markdown files, including agent configs, skills, and conventions, and lets you add them to your vault with provenance recorded. A GitHub search integration picks up anything the curated index misses. Files come from their own repositories; Discover is an index, not a host.