AI harnesses (Claude Code, Cursor, Codex, Qwen Code…)

WriteLikeMe ships a built-in MCP server (writelike-me mcp) and a CLI, so any coding agent can turn your writing into an SVG style code and then write in your voice. Everything runs locally; no API key, no network calls after install.

The workflow:

1. Create a style code once: your writing sample → style/me.svg.
2. Whenever the agent writes prose for you, it reads a ~100-token style brief from that SVG and follows it.
3. Optionally, it checks its draft with compare_styles and revises.

Token use

The tools never return SVG markup to the model. generate_style_svg writes the file and returns only its path and the brief; describe_style returns about seven short lines. On the CLI, always use -o file.svg and --brief.

1. One-time setup in a terminal (outside the TUI)

Requires Node.js 14+ (18+ recommended). npx -y downloads the package on first use, so no install is needed. To pin a version for a project instead:

npm install --save-dev @ai-clyde/writelikeme   # then use: npx writelike-me ...

Create your style code from any text, Markdown or email export you wrote:

npx -y @ai-clyde/writelikeme generate -f my-writing.md -o style/me.svg
npx -y @ai-clyde/writelikeme brief --style-code style/me.svg

Use a few hundred words or more; short samples give weak signals. You can also make the SVG on the website demo and save it as style/me.svg.

Then register the MCP server with your harness below.

2. Register the MCP server

All harnesses run the same command: npx -y @ai-clyde/writelikeme mcp.

Windows

If a client reports that it can't start npx, wrap it with cmd /c: "command": "cmd", "args": ["/c", "npx", "-y", "@ai-clyde/writelikeme", "mcp"]. For Claude Code: claude mcp add writelikeme -- cmd /c npx -y @ai-clyde/writelikeme mcp.

Claude Code

Terminal:

claude mcp add writelikeme -- npx -y @ai-clyde/writelikeme mcp
# everywhere, not just this project:
claude mcp add --scope user writelikeme -- npx -y @ai-clyde/writelikeme mcp
# shared with your team via .mcp.json in the repo:
claude mcp add --scope project writelikeme -- npx -y @ai-clyde/writelikeme mcp

Inside Claude Code:

/mcp

shows writelikeme as connected. To run the CLI without leaving the session, prefix it with !:

! npx -y @ai-clyde/writelikeme brief --style-code style/me.svg

Optional slash command. Save as .claude/commands/write-like-me.md, then use /write-like-me a LinkedIn post about our launch:

---
description: Write in my style using my WriteLikeMe style code
argument-hint: <what to write>
allowed-tools: Bash(npx -y @ai-clyde/writelikeme:*)
---
My style brief:
!`npx -y @ai-clyde/writelikeme brief --style-code style/me.svg`

Write the following in that style. Match the voice, not the topic: $ARGUMENTS

Cursor

Create .cursor/mcp.json in the project, or ~/.cursor/mcp.json for all projects:

{
  "mcpServers": {
    "writelikeme": {
      "command": "npx",
      "args": ["-y", "@ai-clyde/writelikeme", "mcp"]
    }
  }
}

Enable it under Cursor Settings → MCP (Tools & Integrations); a green dot means it's running. The Cursor CLI (cursor-agent) reads the same file; cursor-agent mcp list confirms it.

Optional always-on rule, .cursor/rules/writelikeme.mdc:

---
description: Use my writing style for prose
alwaysApply: false
---
When writing prose for me (docs, emails, posts, release notes), call the
writelikeme `describe_style` tool with style_code_path `style/me.svg` and follow the brief.

OpenAI Codex CLI

Terminal:

codex mcp add writelikeme -- npx -y @ai-clyde/writelikeme mcp

Or add it to ~/.codex/config.toml yourself:

[mcp_servers.writelikeme]
command = "npx"
args = ["-y", "@ai-clyde/writelikeme", "mcp"]

Inside Codex, /mcp lists the server and its tools. Put the project instruction (section 4) in AGENTS.md.

Qwen Code

Terminal:

qwen mcp add writelikeme npx -y @ai-clyde/writelikeme mcp                  # user settings
qwen mcp add --scope project writelikeme npx -y @ai-clyde/writelikeme mcp  # this project only

Or edit .qwen/settings.json (project) or ~/.qwen/settings.json (user):

{
  "mcpServers": {
    "writelikeme": {
      "command": "npx",
      "args": ["-y", "@ai-clyde/writelikeme", "mcp"]
    }
  }
}

Inside Qwen Code, /mcp shows the tools, and ! runs shell commands (! npx -y @ai-clyde/writelikeme brief --style-code style/me.svg). Put the project instruction in QWEN.md.

Gemini CLI

gemini mcp add writelikeme npx -y @ai-clyde/writelikeme mcp

Same mcpServers JSON as Qwen Code, in .gemini/settings.json. Use /mcp inside (or gemini mcp list in a terminal), and put the project instruction in GEMINI.md. Gemini CLI must be signed in or have GEMINI_API_KEY set before gemini mcp commands run.

VS Code (GitHub Copilot agent mode)

.vscode/mcp.json:

{
  "servers": {
    "writelikeme": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@ai-clyde/writelikeme", "mcp"]
    }
  }
}

Click Start above the server entry, then pick the tools in Copilot Chat's tool menu (agent mode).

Windsurf

~/.codeium/windsurf/mcp_config.json, same mcpServers JSON as Cursor. Refresh in Windsurf Settings → Cascade → MCP servers.

OpenCode

opencode.json in the project:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "writelikeme": {
      "type": "local",
      "command": ["npx", "-y", "@ai-clyde/writelikeme", "mcp"],
      "enabled": true
    }
  }
}

Any other MCP client (Cline, Roo Code, Zed, Claude Desktop…)

Use the standard stdio entry: command npx, args ["-y", "@ai-clyde/writelikeme", "mcp"]. If the client has no MCP support, use the CLI (section 5).

3. Prompts to paste into the harness

These work in any harness once the MCP server is connected.

Create my style code

Use the writelikeme generate_style_svg tool with text_path "my-writing.md" and
output_path "style/me.svg". Show me the returned brief only.

Write like me

Call writelikeme describe_style with style_code_path "style/me.svg", then write
<WHAT YOU WANT> following that brief. Match my voice, not my topics.

Write, check and revise once

Write <WHAT YOU WANT> in my style (writelikeme describe_style, style_code_path
"style/me.svg"). Then run compare_styles with reference_path "my-writing.md" and
your draft as candidate. If similarity is below 0.85, revise once using the
largestDifferences, and tell me the final score.

Check a draft in CI

Use the writelikeme check_draft tool with draft_path "<DRAFT.md>",
style_code_path "style/me.svg" and min 0.75. Report the score and any fixes.

Equivalent CLI step for CI:

npx -y @ai-clyde/writelikeme check --style-code style/me.svg -f README.md --min 0.75

Exit code 1 when the score is below --min, so it fails the build on drift.

Restyle existing text

Rewrite docs/announcement.md in my style using writelikeme describe_style
(style_code_path "style/me.svg"). Keep every fact and link; change only the voice.

No MCP? Use the CLI from the agent

Run `npx -y @ai-clyde/writelikeme brief --style-code style/me.svg` and follow that
brief to write <WHAT YOU WANT>. Don't print the SVG file.

Aider has no MCP, but /run adds command output to the chat:

/run npx -y @ai-clyde/writelikeme brief --style-code style/me.svg

4. Project instruction (CLAUDE.md, AGENTS.md, QWEN.md, GEMINI.md)

Add this so the agent applies your style without being asked each time:

## Writing style
When writing prose for me (docs, READMEs, emails, posts, release notes), first get my
style brief: the writelikeme MCP `describe_style` tool with style_code_path
`style/me.svg`, or `npx -y @ai-clyde/writelikeme brief --style-code style/me.svg`.
Follow it. Never paste the SVG contents into the conversation. This does not apply to
code or code comments.

5. CLI reference

writelike-me analyze  [text] [-f file]... [--brief]       # style scores (JSON)
writelike-me generate [text] [-f file]... [-o out.svg] [--width N] [--height N] [--no-words]
writelike-me brief    --style-code me.svg              # ~100-token style brief
writelike-me inspect  me.svg [--brief]                 # profile stored in an SVG
writelike-me write    --style-code me.svg "prompt"     # template draft, no AI
writelike-me check  --style-code me.svg [draft text | -f draft.md | -] [--min 0.8] [--json]
writelike-me compare  "reference" "candidate" [--brief]
writelike-me compare  --reference-file a.md --candidate-file b.md
writelike-me mcp                                       # MCP server over stdio
writelike-me --version

Use - as the text to read stdin: cat notes.md | writelike-me analyze - --brief. Without -o, generate prints the SVG to stdout. Run any of these with npx -y @ai-clyde/writelikeme <command> if the package isn't installed.

MCP tools

Tool	Arguments	Returns
generate_style_svg	text, text_path or text_paths, output_path, width, height, include_words	file path, size, brief
describe_style	style_code_path, svg, text or text_paths	style brief (plain text)
analyze_text	text, text_path or text_paths, full	compact summary (full JSON if full)
read_style_code	style_code_path or svg, full	compact summary of an SVG's profile
compare_styles	reference/reference_path, candidate/candidate_path	similarity + 6 largest differences
write_like_me	prompt, style_code_path or svg	template-based local draft
check_draft	draft or draft_path, style_code_path or svg, min	score, pass flag, up to five issues with fixes

Relative paths resolve against the directory the harness starts the server in (normally the project root).

MCP prompts (slash commands)

The server also advertises two MCP prompts, which clients such as Claude Code expose as slash commands: /mcp__writelikeme__write_like_me and /mcp__writelikeme__check_my_draft.

Prompt	Arguments	What it returns
write_like_me	task (required), context (optional)	The style brief from style/<context>.svg or style/me.svg, followed by an instruction to write <task> in that voice itself (without calling the template write_like_me tool). When no style code exists, it tells the model to call generate_style_svg on a writing sample first, then describe_style.
check_my_draft	draft (required), context (optional)	Asks the model to call check_draft on the draft against style/<context>.svg or style/me.svg (pass threshold 0.8), and revise once using the listed fixes if it does not pass.

Unknown prompt names and missing required arguments return JSON-RPC error -32602.

Troubleshooting

Symptom	Fix
Server shows as failed or disconnected	Run npx -y @ai-clyde/writelikeme --version in a terminal. If that fails, install Node.js 18+ and check that npx is on the PATH the harness sees.
spawn npx ENOENT on Windows	Use the cmd /c form above.
Could not read …/style/me.svg: file not found	Paths are relative to the server's working directory. Use an absolute path or open the harness in the project root.
Unsupported or invalid … version	The SVG wasn't made by WriteLikeMe, or was edited. Regenerate it.
Slow first start	npx downloads the package once. Pin it with npm i -D @ai-clyde/writelikeme and use "command": "node", "args": ["node_modules/@ai-clyde/writelikeme/cli.js", "mcp"].

Verify a checkout of this repository with npm test. It covers the library, the Handler, every CLI command, a real MCP stdio session (all six tools, two prompts, error paths and malformed input), the style-accuracy benchmark, and the contents of the published package.

The server speaks MCP protocol versions 2024-11-05, 2025-03-26 and 2025-06-18 over stdio, and logs nothing to stdout except protocol messages.