Writhink MCP server
Your client comments.
Your agent gets the context.
Writhink is a remote MCP server. Connect Claude Code, Cursor or Claude Desktop to your account and your agent reads the feedback your client left on the page — the comment, the element, the screenshot — then fixes it, replies, and closes the item.
Free while we’re in beta. No credit card, no waitlist.
The loop
From what the client said to code that ships
- 1
Client comments
They open your link — no account — and pin a comment on the exact element.
- 2
Writhink captures
Page, element anchor, device, viewport and a screenshot of the spot, attached to the comment.
- 3
Agent reads
get_feedback hands your agent the whole thing as structured context, not a paraphrase.
- 4
Agent implements
It edits the code in your repo, the way it already does with any other task.
- 5
Loop closes
reply_to_feedback answers the client in the thread; set_feedback_status marks it resolved.
Connect it
One authorization, then it works
You sign in once with your Writhink account and approve the connection. There is no API key to generate, paste or rotate — and nothing to put in a config file that could leak.
Claude Code
One command in your terminal.
claude mcp add --transport http writhink https://mcp.writhink.com/mcpCursor
Settings → MCP → add a server with this URL, transport HTTP.
https://mcp.writhink.com/mcpClaude Desktop & claude.ai
Settings → Connectors → add a custom connector with this URL.
https://mcp.writhink.com/mcpWhat the agent can do
Seven tools, and no others
This is the complete surface. An agent connected to Writhink can call these and nothing else — it cannot read your account settings, your team, your billing or anyone else’s organization.
| Tool | What it does |
|---|---|
list_projects | Lists the projects your account can see, with the count of open feedback items |
get_feedback | Reads the comments: text, page URL, pinned element, device, viewport, screenshot URL, status, thread |
get_share_link | Returns the review link for a project, ready to send |
set_feedback_status | Moves an item between open, in progress, in review and resolved |
reply_to_feedback | Posts a reply the client sees in the public thread |
create_project | Creates a project from a URL, so a review link exists |
setup_widget | Sets up the embeddable feedback widget for a site |
How access works
The agent is you, with your permissions
OAuth 2.1, no keys
You approve the connection in a browser, like any other app. Nothing secret ends up in a config file on disk.
Your identity, not ours
Every query the server makes runs as your user. There is no service key that could see more than you can.
The database decides
Row-level security is the boundary, not application code. A project you cannot open, the agent cannot read.
Revocable, instantly
Revoke the authorization from your profile whenever you want. Access stops there and then.
One current limitation, stated plainly: an agent works within your active organization. If you run more than one, switch in Writhink and reconnect to work on another.
FAQ
Before you connect it
- Do I need a paid plan?
- No. The MCP server is available to every account, and everything is free while Writhink is in beta. You do not need a credit card to connect it.
- What can the agent actually see?
- Exactly what your own account can see, and nothing else. The connection authenticates as you over OAuth 2.1, and every query the server makes runs under your identity with the database's row-level security applied. There is no service key and no shared secret that could widen that.
- Can it change things without asking me?
- It can do anything the seven tools allow — including replying to a client and resolving items — because that is the point of connecting it. Your AI client is where you control that: both Claude Code and Cursor ask before running a tool unless you tell them otherwise. Treat replying to a client as an action worth approving.
- How do I disconnect it?
- Revoke the authorization from your Writhink profile at any time. The agent loses access immediately; nothing else about your account changes.
- It only sees one of my organizations.
- That is a real current limitation, not a bug on your side: the agent works within your active organization. Switch organization in Writhink and reconnect to work on another one.
- Which AI clients work?
- Any client that speaks remote HTTP MCP. We have used it with Claude Code, Cursor and Claude Desktop. Others that support custom MCP connectors should work the same way — the server does not special-case any client.
Start
Give your agent the client’s words
Create a project, send the link, let your client comment. Then ask your agent what they asked for.
New to Writhink? see how the feedback part works or read the Claude Code walkthrough.