The agent-native document editor. Humans and AI agents collaborate in the same markdown document — with real-time editing, comments, suggestions, and full authorship tracking.
Try Comment.io → · Agent start · REST reference · Optional skill
If Comment.io tools are already available, follow their own workflow immediately.
For an existing comm, the standard MCP surface uses open_comm, then
read_comm; use create_comm only when the human requested a new comm, never as
setup verification. An account-scoped document request tool follows its own
relative-path contract. For a supplied comm plus authenticated HTTPS, resolve a
clean shortlink once with Accept: text/html, application/json;q=0, without Authorization or automatic redirects, and accept
only an exact token-bearing Comment.io /d/{slug} Location. Then send
X-Comment-Share-Ingress: 1 on the first authenticated ?docs GET at that
final origin, switch to returned your_token, and follow that handoff. With URL
fetch only (no browser UI/headers), use any supplied share URL now. Add
mode=agent without removing its query and read the first response directly.
For a bare slug or token-free comm URL, first try
the final Comment.io origin's /d/{slug}?mode=agent; continue only when markdown is
non-null. If it is null or says no readable state, ask once for Share → Copy
for agent. The envelope is read-only. With interactive browser control, use
the visible editor and comments UI. A hosted MCP connector takes a slug,
token-free document URL, or exact clean CMNT/configured shortlink only in
url_or_slug; never a raw token-query URL or clean link embedded elsewhere. A
local MCP tool may accept the full share URL only when its description says so.
Use the startup index
only when the current route does not cover the task.
When the user asks for a new comm and no existing tool or browser route can create it, follow the live REST setup path. It gives agents one current route through identity, anonymous fallback, per-doc tokens, creation, and revision-safe edits. Endpoint shapes and recovery behavior live in the exact REST reference. There is no SDK requirement; the protocol is plain HTTPS.
| Comment.io | Google Docs | Notion | |
|---|---|---|---|
| Agent REST API | ✅ Full CRUD + comments + suggestions | ❌ Batch import/export only | ❌ No document editing API |
| Agent identity | ✅ Registered handles + doc-scoped tokens | ❌ | ❌ |
| Provenance | ✅ Per-edit attribution, human vs AI | ❌ | ❌ |
| Multi-agent | ✅ Multiple agents, real-time, with loop prevention | ❌ | ❌ |
| @mention agents | ✅ Connector inbox, local daemon, or webhooks | ❌ | ❌ |
| No login required | ✅ | ❌ Google account | ❌ Account required |
| Suggestion mode | ✅ API + UI | ✅ UI only | ❌ |
| Real-time sync | ✅ Yjs CRDTs + WebSocket | ✅ OT | ✅ |
- Claude Code — Plugin with Comment.io skills and local CLI notification checks
- ChatGPT / Claude chat — Hosted MCP connector with OAuth; no local install required
- OpenClaw — Channel plugin with an account-scoped
comment_io_requesttool after exact binding; automatic @mention delivery is separate and requires the matching installed full-handle profile plus a persistent Comment CLI/daemon on a long-lived computer (focused setup) - Codex — Universal skill installer plus optional local MCP through the Comment.io CLI
- Any HTTP client — It's REST. If you can
curl, you can collaborate.
See the integrations/ directory for setup guides.
- CLI on npm —
@comment-io/cli:npm install -g @comment-io/cli - Agent/machine docs — comment.io/llms.txt (compact startup index) · exact REST reference
- Engineering-workflow skills — comment-hq/skills:
npx skills add comment-hq/skills(skills.sh) - Claude Code plugin — comment-hq/comment-io-claude-code-plugin
- OpenClaw plugin — comment-hq/openclaw-plugin
CommentFS can project the configured Comment.io library scope into local Markdown files for search, context, and agent inspection. Follow the live local-sync guide; it starts by selecting the exact saved origin/account and letting the registry resolve its scoped home, then covers status, login, one-shot sync, and the optional background worker without relying on an ambient account.
Projections are read-only by default. The guide also covers the separate human browser consent that can make My Files projections writable on that computer. Shared With Me and Team Wiki projections remain read-only.
| Agent start | Capability-first startup index |
| API Reference | Exact REST endpoint behavior and recovery |
| llms.txt | Compact machine-readable startup index |
| comment.SKILL.md | Short skill file that points agents to llms.txt |
| OpenClaw skill | OpenClaw-specific skill stub |
| agent.json | .well-known/agent.json for agent discovery |
| What is agent-native editing? | The concept explained |
- 🐛 Issues — bug reports and feature requests
- 💬 Discussions — questions, ideas, show & tell
MIT — see LICENSE.