amplifiedlabs

presend

A “should I send this?” checker, shipped as an MCP server.

Paste a draft email / LinkedIn post / proposal / feedback note. presend grades it 0–10, flags risky phrasing (implied commitments, vague timeframes, off-tone, bias, oversharing…), and — on request — rewrites it. Optional per-recipient profiles let you grade against a real person’s wiring instead of a generic ideal.

What it does

Five MCP tools:

Tool	What it does
check_message	Score a draft. Returns a markdown table of red/yellow/green flags + one-line summary.
fix_message	Rewrite the draft applying the flags from check_message.
save_profile	Save a recipient profile (boss, client, partner) and their communication preferences.
list_profiles	Show saved profiles.
delete_profile	Remove a saved profile by id.

Modes: email, linkedin, proposal, feedback, general. Each mode applies its own checklist (e.g. email mode looks for implied commitments and vague timeframes; feedback mode looks for bias and unactionable critique; proposal mode looks for risk language and missing sections).

Profiles live at ~/.presend/profiles.json. A profile carries a free-text description of how a specific recipient prefers to be communicated with — direct/diplomatic, bullets/prose, formal/casual, sensitivities — plus a default language. Pass profile_id to check_message/fix_message to grade against that person rather than a generic ideal.

Languages. presend auto-detects the language of your draft and responds in that language. Mixed-language drafts (e.g. PT + EN) get flagged.

How it works

presend is an MCP server, but it does not call an LLM itself. Each tool returns a structured instruction block that the host LLM (the model you’re talking to) executes in its very next turn. No API keys, no sampling, no outbound network calls from the server. The MCP client doesn’t even need to support sampling/createMessage.

This means presend works with any MCP-compatible host — Claude Code, Claude Desktop, Cursor, Cline, etc. — as long as the host model is capable of following instructions and rendering markdown.

Prerequisites

• Node.js 18+
• An MCP-compatible client (Claude Code, Claude Desktop, Cursor, Cline, …)

Install

Option A — Claude Code (CLI), via claude mcp add

The simplest path on Claude Code. From anywhere:

git clone https://github.com/PTthe13/amplifiedlabs.git
cd amplifiedlabs/presend
npm install
npm run build

claude mcp add presend node "$(pwd)/dist/index.js"

Restart your Claude Code session. The five mcp__presend__* tools appear.

Option B — Claude Code plugin (local marketplace)

This folder ships with .claude-plugin/plugin.json and .mcp.json, so it can be added as a local marketplace and installed by name:

# from inside Claude Code
/plugin marketplace add /absolute/path/to/amplifiedlabs
/plugin install presend

Or install directly from the folder:

/plugin install /absolute/path/to/amplifiedlabs/presend

Option C — Claude Desktop config

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or the equivalent on your platform:

{
  "mcpServers": {
    "presend": {
      "command": "node",
      "args": ["/absolute/path/to/amplifiedlabs/presend/dist/index.js"]
    }
  }
}

Restart Claude Desktop.

Uninstall

Option A — Claude Code CLI

claude mcp remove presend

Option B — Claude Code plugin

/plugin uninstall presend
# optional: drop the local marketplace too
/plugin marketplace remove amplifiedlabs

Option C — Claude Desktop

Remove the presend entry from claude_desktop_config.json and restart.

Profiles cleanup (optional)

Profiles live outside the repo and survive uninstall. To wipe them:

rm -rf ~/.presend

Usage

Once installed, just talk to your assistant:

“Check this email before I send it. Mode: email. Profile: boss-joao.”

Olá João, vou tratar do relatório Q3 em breve, não te preocupes.

Your assistant calls check_message and renders:

SCORE: 4.5/10 — Mode: Email → João

Level	Flag	Detail
🔴	DISMISSIVE_TONE	“não te preocupes” pode soar condescendente para um chefe stressado
🔴	IMPLIED_COMMITMENT	“vou tratar” é promessa firme sem âmbito
🟡	AMBIGUITY	“em breve” não dá prazo
✅	BREVITY	curto, alinha com a preferência do João

Summary: Tom dispensa-te, prazo vago — define data e remove a tranquilização.

💡 Want a corrected version? Ask: fix_message

Then:

“Apply those flags.”

Your assistant calls fix_message and prints the rewrite.

Profile lifecycle example

“Save a presend profile: id boss-joao, name ‘João (chefe)’, direct/terse style, prefers bullets, dislikes long emails, sensitive to tone when stressed, language pt.” “List my presend profiles.” “Delete the boss-joao profile.”

Security

• No API keys handled by the server. No outbound network calls. Analysis runs inside the host LLM you’re already talking to.
• Profiles are stored locally at ~/.presend/profiles.json. They never leave your machine except as part of the prompt your assistant generates.
• .env / .env.* are listed in .claudeignore and .gitignore as a precaution. The server reads no env vars.

Development

npm install
npm run build
node dist/index.js   # stdio MCP server, talks JSON-RPC over stdin/stdout

To iterate locally:

npm run dev          # tsc --watch
pkill -f "presend/dist/index.js"   # force MCP client to respawn with new dist

Stack

TypeScript · @modelcontextprotocol/sdk · stdio transport · zero deps beyond MCP.

Status

🧪 Lab experiment. Used internally before sending tricky replies. PRs welcome for new modes and check categories.

License

MIT.

This site is open source. Improve this page.