linkedin-formatter-mcp

An MCP server that lets an AI assistant style LinkedIn posts with real Unicode.

The problem it solves

LinkedIn has no formatting controls. Every "bold" post is a Unicode substitution, and Markdown does nothing: **bold** pastes in as literal asterisks.

Models are unreliable at this. Script capital B is U+212C, in Letterlike Symbols, not one past script A, because the Mathematical Alphanumeric block has holes where Unicode had already assigned those letters decades earlier. Four ranges need exception tables. A model working from memory gets this subtly wrong, and the failure renders as an empty box on someone's phone.

Models also cannot count LinkedIn posts correctly, because LinkedIn counts UTF-16 code units and a styled character costs two.

This server replaces guessing with a function call.

Setup

Claude Code, in one line:

claude mcp add -s user linkedin-formatter -- npx -y linkedin-formatter-mcp

Every other client takes the same thing as configuration:

{
  "mcpServers": {
    "linkedin-formatter": {
      "command": "npx",
      "args": ["-y", "linkedin-formatter-mcp"]
    }
  }
}

npx fetches the server on demand, so there is nothing to install first. To pin it locally instead:

npm install -g linkedin-formatter-mcp

then use "command": "linkedin-formatter-mcp" with no args.

Where the configuration goes (Node 18 or newer):

• Claude Desktop: claude_desktop_config.json (Settings, then Developer, then Edit Config)
• Cursor: ~/.cursor/mcp.json
• Windsurf: ~/.codeium/windsurf/mcp_config.json
• Codex: an [mcp_servers.linkedin-formatter] block with command = "npx" and args = ["-y", "linkedin-formatter-mcp"]

Restart the client after registering. MCP servers attach at session start, so a server added to a running session stays invisible until then. If you need the same conversion right now, the linkedin-fmt CLI is the identical core and needs no restart: npx -y linkedin-fmt style "Shipped" --bold.

Once registered and reloaded, the tool descriptions reach the model automatically at session start, so it learns the .linkedin convention without being told.

Tools

| Tool | Purpose | |---|---| | apply_family | Style text as a font family, with optional bold/italic. The main one. | | apply_style | Style text by an exact style id | | apply_mark | Layer strikethrough or underline | | strip_formatting | Convert styled text back to plain ASCII | | count_characters | Count against the 3,000 limit, in UTF-16 units | | list_families | The 11 families and which axes each supports | | list_styles | Every letterform style and combining mark id | | tier_samples | One hook line rendered in the three tiers, one line each, to show before asking |

All eight are read-only, non-destructive, closed-world pure functions, and are annotated as such so clients need not prompt before running them.

Guarantees

No network. stdio transport only. The server opens no socket, makes no request, and has no telemetry. HTTP was deliberately not implemented: a listening socket, even on localhost, would need its own threat model against a tool whose point is that nothing leaves the machine.

No filesystem access. It transforms text handed to it. It does not read or write files, including the .linkedin files it talks about.

Fail closed. A character with no equivalent in the target style is left unchanged rather than swapped for a lookalike.

Honest results. Unicode is sparse: monospace has no bold, script has no italic. Asking for an unavailable axis drops it and says so in the response, so the model does not report bold text that is not bold.

Lossless round trip. strip_formatting reverses styling exactly, for every family, asserted over the full A-Za-z0-9 range.

Working with the extension

The companion VS Code extension renders any .linkedin file as a live LinkedIn card beside the editor, so a human can see and adjust what the model produced:

code --install-extension kaushiknaru.linkedin-formatter

Cursor and Windsurf take the same id from Open VSX, with cursor or windsurf in place of code.

Because the formatting is the text, both sides edit the same file with no sync step and no markup layer to reconcile.

A note for agents sharing a file with a human: re-read before writing. The user may have edited it since you last looked.

Related

• VS Code extension: LinkedIn Formatter
• CLI: linkedin-fmt, the same core for shell and scripts

All three share one conversion core with no duplicated logic.

Licence

MIT