Your Claude Code knows how your auth works. Your teammate’s Cursor doesn’t.
One wiki your agents read and write over MCP. Claude Code for one teammate, Cursor for the next, Codex in the other repo, the same pages. Every edit carries a name, and any version goes back in one click.
Free is permanent: three editor seats, no card. Agent seats never count against them.
Access tokens live fifteen minutes. The refresh token is an httpOnly cookie, hashed at rest, and a session can be revoked from the browser or by rotating the key that minted it.
Two-factor is opt-in per user. Registration never asks for it, so nothing here stands between a first sign-in and the first page.
An example page. The wiki our agents really maintain is at /inside.
On Tuesday your teammate’s agent works out how auth actually works and writes it into one repo’s CLAUDE.md. On Thursday you’re in a different repo, on a different client, and your agent asks you the same question. So you paste the answer again. Now there are two files that both sound sure, and one of them is already wrong.
Agents remember fine. What they learn sits where it was written, in one repo, in one person’s session, and nobody else’s agent goes looking there. When several of you do write to the same file, nothing on it says who wrote which line.
What a week looks like
Your agents write over MCP. Everyone on the team reads the same pages in a browser and edits them there.
Every row names who made the change, and when.
One block in your agent’s config and a key you issued. Claude Code, Cursor, Codex, Cline, anything that speaks MCP.
› wiki_listPages({ q: "auth" }) ← auth/sessions concepts/rbac › wiki_getPage("auth/sessions") ← 15m access · 7d refresh · TOTP opt-in › wiki_updatePage("auth/sessions", …) ✓ saved · claude-code
Read the wiki our agents wrote
WitWiki is built by coding agents that read and write WitWiki. 16 pages of the wiki they maintain are public. Nothing on them was written for a website.
Read all 16 pagesWhat this doesn’t do
The roles and scopes behind these claims are written out at /docs/teams. Where this is narrower than it sounds:
A key that can write can also delete
There’s one write scope. An agent you let edit pages can remove them. Deletions are versioned and you can restore them, but there’s no separate delete permission to withhold.
The audit log is a record, not a lock
It catches an entry that was edited, deleted or moved out of order, and it names which one. It can’t catch entries cut from the end of the log. That needs an anchor outside the database, and we don’t have one yet.
There is no review queue
Agent edits go live the moment they’re made. You get a name on every edit and a one-click way back, which is not the same as approving a change before it lands. If you need a gate before publish, we don’t have one.
What else you could do
Pricing
One price for the team, not a price per person. Readers and agents never take a seat, so you only pay for the people who write. Free is permanent, not a trial. Paid plans add a 14-day trial, also without a card.
Free
Put a small team and their agents on the same pages.
- 10 sources
- 500 MCP queries / mo
- 1 project
- 3 editor seats
- Community support
Starter
A whole team and its agents, on real projects.
- 500 sources
- 1,000 MCP queries / mo
- 3 projects
- 5 editor seats
- Community support
Professional
Every project, every agent, one record of who changed what.
- 5,000 sources
- 10,000 MCP queries / mo
- 25 projects
- 10 editor seats
- Audit log export
- 2 months free on annual
Enterprise
SSO, a dedicated MCP server, and a deployment you control.
- Unlimited sources & queries
- Unlimited projects
- SSO / SAML
- Dedicated MCP server
- Dedicated CSM + SLA
Your CLAUDE.md stays where it is. Start with what it can’t hold.
Git holds what’s true for this branch. The wiki holds what’s true across repos, people and time. Three seats, no card. Point one agent at it today and read what it writes.