Plug Plain into Claude Code in 30 seconds

You're in Claude Code, mid-task. You've finished refactoring the payments module. The PR is up. Your manager asks: "Can you put this in a one-pager for the team meeting tomorrow?"

The traditional flow: leave the terminal, open a doc tool, paste the diff, ask AI to summarize, copy back, fix formatting, share a link, return to terminal. Twenty minutes minimum.

The Plain MCP flow: stay in the terminal. Tell Claude Code: "Write a one-pager about the payments refactor and give me the link." Done in thirty seconds.

Setup, once

# Install the CLI
pnpm dlx @plain/cli login

# Start the MCP server
plain mcp

That's it for the Plain side. Now wire it into Claude Code.

Edit ~/.config/claude-code/config.json:

{
  "mcpServers": {
    "plain": {
      "command": "plain",
      "args": ["mcp"]
    }
  }
}

Restart Claude Code. The plain.* tools now appear in its tool listing — generate, edit, share, export.

What it can do

Once connected, Claude Code can:

• Generate a deck from a research summary it just wrote. It calls plain.generate_deck with the context. You get a share link in seconds.
• Write a doc from a PR diff or design notes. It calls plain.generate_doc. The output is a real web page, not a markdown blob in your terminal.
• Build a dashboard from a CSV in your repo. It calls plain.generate_sheet with the data and a hint about what story to tell.
• Edit any of them by ID. "Make slide 3 shorter, and add a closing." Claude reads the source, sends the edit instruction, and Plain applies it.
• Export to .pptx / .docx / .xlsx when the downstream consumer wants a file, not a link.

Why this matters

AI coding agents have great reasoning but no fingers. They can think through a refactor; they can't open Keynote. MCP closes that gap for a specific shape of work: anything that ends in "produce a deck / doc / dashboard."

That's most engineering-adjacent work. Release notes. Design reviews. Incident write-ups. Architecture diagrams. Investor updates. Numbers dashboards. The agent does the thinking; Plain does the producing.

Sandboxing for shared sessions

If you don't want the agent to spend credits or mint public links unchecked, you can constrain it:

plain mcp --readonly                # no generate / edit
plain mcp --no-share                # no public share links
plain mcp --budget=50               # cap credits per session
plain mcp --allow=generate_deck     # whitelist

Useful when you're running a Codex agent unattended or sharing a terminal with a colleague.

It works in Cursor too

Cursor speaks MCP. So does Codex CLI. So does any tool that supports the protocol. The same plain mcp command works for all of them — you just configure each tool to know about it.

See the MCP server docs for the full setup per tool.

The deeper point

We didn't build Plain as another standalone app you have to remember exists. We built it as a service your existing AI agents can call. The terminal is fine. Cursor is fine. Claude Code is fine. Just give them a way to ship the artifact at the end of the task.

That's MCP. And that's why Plain ships with a first-class server.