What is MCP?
MCP (Model Context Protocol) is an open standard for connecting AI applications to external systems. It lets an AI agent reach out to data sources, tools and workflows outside of its own chat window. So instead of you copying and pasting feedback into a chat, the agent goes and gets it.
Pastel runs an MCP server at https://api.usepastel.com/mcp. Any AI agent that supports MCP can connect to it, which means Claude, ChatGPT, Cursor and others all work with Pastel without needing a separate integration for each one.
You can read more about the standard itself at modelcontextprotocol.io.
What you can do with it
Once your agent is connected, you can ask it to work with your Pastel canvases and comments in plain language. For example:
"List the canvases in my Website Redesign project."
"Summarize the unresolved comments on the pricing page canvas, grouped by page."
"Take comment #14, work out what's causing it in this codebase, fix it, reply explaining the fix, and resolve it."
"Assign every comment about copy to Priya and mark them in progress."
"Create a website canvas for staging.example.com and invite the design team to leave feedback."
The real advantage shows up when your agent also has access to your codebase - in Cursor, Claude Code or Codex, for example. Then it can go from a comment on a canvas all the way through to a fix in your code and a reply on the comment, in a single pass.
The built-in investigation guide
Reading a comment is only half the job. The harder part is working out what the comment is actually pointing at.
Pastel handles that for you. Whenever an agent views a comment, we include step-by-step instructions on how to track down what that comment refers to, tailored to the type of canvas it's on:
Website canvases: which page to open, which element on that page the comment is pinned to, and where to find the screenshot.
PDF canvases: which page, and the position on it.
Video canvases: the timestamp.
From there, the guide tells the agent to find the root cause in your codebase, fix it, reply explaining what it changed, and mark the comment as resolved. You don't need to write any of this into your prompt - it comes along with every comment automatically.
Permissions and access
Your agent acts as you. It can see and change exactly the canvases and comments that you can, and nothing more.
When you connect an agent, you sign in to Pastel from that agent and approve access, the same way you would connect any other app to your account.
Every connected agent is listed in the MCP section of your profile settings, and you can disconnect any of them there at any time. Disconnecting takes effect immediately - the agent loses its access to your Pastel account until you connect it again.
Setting it up
Setup takes a minute or two. Pick your agent below:
Connecting Pastel to Claude (Claude.ai, Claude Desktop and Claude Code)
Connecting Pastel to ChatGPT and Codex (ChatGPT.com, ChatGPT Desktop and the Codex CLI)
You can connect as many agents as you like, and the same instructions are available in the MCP section of your profile settings.
Tool reference
The server exposes 19 tools. Your agent picks the right ones on its own based on what you ask it for. This list is here so you know what's possible and what isn't.
Read tools
Tool | What it does |
| List canvases you have access to |
| Fetch one canvas and its assets |
| List the comments on a canvas |
| Fetch one comment in full, with replies and metadata |
| List who can be @mentioned on a canvas |
| List your team's comment labels |
| View an image attached to a comment or reply |
| View the screenshots captured with a comment |
Write tools
Tool | What it does |
| Create a canvas |
| Update a canvas' title, URL, password or visibility |
| Email people an invitation to leave feedback |
| Reply to a comment |
| Edit one of the agent's own replies |
| Resolve a comment, or move it through the statuses |
| Assign a comment to a teammate, or unassign it |
| Step 1 of uploading a canvas asset |
| Step 2 of uploading a canvas asset |
| Step 1 of attaching a file to a reply |
| Step 2 of attaching a file to a reply |
Things worth knowing
Filtering is built in. You don't need to ask for everything and sort through it yourself. Comments can be filtered by status, assignee, author, page URL, label, privacy and date, and searched by their text. Canvases can be filtered by project, type, owner and date, and searched by title or URL.
Agents can only edit their own replies. An agent can reply to any comment you can see, and it can edit replies that it wrote itself. It can't edit replies written by you or your teammates.
Sharing sends real email. share_canvas sends an actual invitation to the addresses given, so agents are instructed to confirm the recipient list with you first. Recipients without a Pastel account get guest access to that canvas.
Guests can be mentioned but not assigned. Only members of your team can have comments assigned to them.
Uploads are limited to 110MB per file, whether it's a canvas asset or a reply attachment.
Videos take time to process. After a video is uploaded it encodes in the background, and won't play until that finishes. Encoding takes roughly as long as the video itself, so a five minute video takes around five minutes. Your agent knows to wait and check back.
Changing a canvas' URL rewrites its comments. Pointing a website canvas at a new URL re-verifies the page and updates the URL on every existing comment, so it's worth only doing when you really do want the canvas pointed somewhere else.
Something not working?
MCP support is new and AI agents are changing quickly, so instructions occasionally go out of date. You may also be using an agent we haven't written instructions for yet.
If something isn't working, or a step doesn't match what you're seeing on screen, let us know and we'll get it sorted.