The biggest thing that shipped this week isn't a component or a tool — it's that your coding agent can now write your docs directly. Point Claude Code, Cursor, or any MCP client at https://mcp.dokly.co/mcp, and it can read your pages, draft new ones, and propose edits to live pages without you ever opening a different app.
If you missed the full launch post, that's the one to read — it covers setup, the full tool list, and exactly what stays a draft until you approve it. This post covers what came right after: two real bugs we found and fixed in the days since it shipped.
A page your agent wrote couldn't be opened safely — now it can#
Here's the bug, plainly: if your agent wrote or edited a page over MCP, and you then opened that same page in the dashboard editor, the editor would quietly re-parse and re-save the content. Sometimes that dropped pieces of what the agent had written. Every single time, it flagged the page as "Unsaved" even though you hadn't touched it.
That's exactly backwards from what "review before it goes live" is supposed to mean — you should be able to open an agent's draft, read it, and trust that looking at it doesn't change it. The editor now carries the content through losslessly, and a page an agent wrote only shows "Unsaved" once you've actually edited it.
One misbehaving connection couldn't be allowed to slow everyone else down#
The MCP server checks who's calling it on every single request. That's normally cheap — until something goes wrong. Earlier this week, one agent client stuck in a reconnect loop turned into roughly 330,000 of those checks in a single day, all from one connection.
The server now caches a short-lived "yes, this credential is valid" answer instead of asking the database every time, and adds rate limits on top. A single misbehaving client can no longer run up backend load for everyone else sharing the same database — not something you'd notice directly, but it's the kind of thing that matters more as more agents stay connected around the clock.
What's next#
With content, structure, and brand now things your agent can already read, write, and adjust through MCP, the next piece is giving it visibility into whether a page is actually still accurate — catching docs that have quietly gone stale against the code they describe, not just checking that they render.
Something you keep working around? Reply to this post or use the contact page — it reaches me directly.
Changelog entries#
Agent-written pages · Fixed · Opening a page written by a connected agent no longer drops content or shows a false "Unsaved" state. MCP server · Changed · Added rate limits and a short credential-check cache so one misbehaving connection can't slow down requests from everyone else.