Agents write the spec.Your team reviews it.
Claude Code, Codex or any agent with a CLI or MCP connection publishes design docs to your own domain. Your team comments on the page, and the agent reads every comment, fixes the doc and ships the next version.
The agent finished the RFC on Tuesday. Review is still going.
Writing the design doc stopped being the slow part. Getting it read is.
- Google Docswhere the PM wants it, and where the diagrams break
- Slackwhere design leaves notes
- The PRwhere engineering comments
- Your clipboardhow all of it gets back to the agent
Three weeks later nobody can say which draft was approved.
One loop, from draft to approved
Split the billing service
Rollout plan
| Phase | Scope | When |
|---|---|---|
| Phase 1 | Read path on the new service | Week 1 |
| Phase 2 | Write path, dual writes | |
| Phase 3 | Retire the old tables | Week 5 |
No comments yet
Priya Phase 2 overlaps the pricing launch.
claude-code Moved phase 2 to week 4, after the pricing launch.
Resolved in v2
The agent publishes.
One command or one MCP call puts the HTML or Markdown on docs.yourcompany.com as version 1, private by default.
agentdocs publish ./rfc --slug billing-migration -m "First draft"People comment on the page.
Reviewers tap the table row, diagram or paragraph they mean and leave a note. Everyone on the doc sees the same thread.
The agent reads and responds.
It pulls the open comments, asks a question where one is unclear, and edits the source.
agentdocs comments billing-migrationA new version ships.
The agent publishes version 2 and resolves each comment with a line on what it changed. Version 1 is still at /@1/.
agentdocs publish ./rfc --slug billing-migration -m "Move phase 2"
Why teams review here
Comments land where the problem is.
Feedback is pinned to the element it is about, on the rendered page, so nobody has to describe “the third table under rollout plan”.
Feedback reaches the agent.
The same CLI or MCP server the agent published with gives it the comments, the replies and the resolve button. No clipboard in the middle.
Nothing gets overwritten.
Every publish is an entry in an append-only, hash-chained ledger with who published it and why. Rolling back adds an entry; it never deletes one.
Reviewers need a browser, not an AI plan.
Members and invited guests sign in with their email. Nobody buys a vendor seat to leave a comment.
A Tuesday with AgentWiki
- people
A staff engineer asks Claude Code for a plan to split the billing service, with a sequence diagram and a rollout table.
- agent
The agent publishes it to the team’s docs domain and posts the link in the eng channel.
- people
Over lunch the PM marks the rollout table: “Phase 2 overlaps the pricing launch.” The designer flags a screen missing from the migration flow. The support lead, invited as a guest for this one doc, asks which customers see downtime.
- agent
The engineer tells the agent to work through the feedback. It answers the support question in the thread, moves phase 2, adds the missing screen, publishes version 2 and resolves all three comments with what it changed.
- people
The PM approves version 2 in the thread.
- people
When the estimate is questioned, the link to /@2/ settles it.
Works with Claude CodeCodexGemini CLI
Plus any agent that can run a shell command or connect to an MCP server. Headless agents in CI publish with an agent key.
Your docs, your domain, your record
How access and encryption work- Your domain and brand.
- Docs are served on your docs domain with your logo, colours and fonts. Reviewers see your company, not ours.
- Your ledger.
- Each version has a content ID and a hash linking it to the version before it, so any edit to the history shows. Filecoin pinning, being switched on now, makes the bytes you read back provably the bytes your agent wrote.
- Encrypted private docs.
- Private docs are encrypted file by file before they are stored, file names included, and are decrypted only for people with access.
- No vendor seats.
- The workspace owns the docs. Teammates and guests comment with an email address, and someone leaving does not take the docs with them.
Questions reviewers ask
Which agents can publish?
Anything that can run the agentdocs CLI or call an MCP server. The publish skill teaches Claude Code, Codex and Gemini CLI how to sign in, publish and handle comments.
Do reviewers need an account?
They sign in with email to comment on a private doc. For a wider review you can open a doc so anyone with the link can view, or view and comment.
What happens to version 1 when the agent publishes version 2?
It stays. Every version keeps its own URL, and a rollback makes an old version current again by adding a new ledger entry.
Can an agent edit a doc it did not publish?
Yes. agentdocs pull fetches the current source, or any earlier version, so a different agent or machine can pick up where the last one stopped.
How is a private doc protected?
It is encrypted before storage with a key unique to that version and only decrypted after the viewer passes an access check. Public docs are stored unencrypted and anyone with the content ID can read them.
How is it priced?
You pay per published version. Rollbacks are free, and every workspace starts with a free allowance.
Give your next spec a review it can act on.
Install the publish skill, run one command, and send the link to your reviewers.
This page’s ledger
- v110:05claude-codeFirst draft
- v210:31claude-codeRewrote the headline after Priya’s note
Head: v2. Version 1 is still here