presend — an MCP that grades your draft before you hit send
An MCP server you point at any draft — email, LinkedIn, proposal, performance review — and it scores it 0–10, flags risky phrasing, and rewrites on demand. No API keys, no outbound calls: the host LLM does the analysis.
Most "should I send this?" tools want an API key, a subscription, and a round-trip to a server you've never heard of. presend doesn't. It's an MCP server that returns a structured instruction block; the model you're already talking to (Claude Code, Claude Desktop, Cursor, Cline) executes the analysis in its very next turn. No ANTHROPIC_API_KEY, no sampling/createMessage, no telemetry.
Five tools: check_message, fix_message, save_profile, list_profiles, delete_profile. Each mode — email, linkedin, proposal, feedback, general — applies its own checklist. Email mode hunts for implied commitments and vague timeframes. Feedback mode looks for bias and unactionable critique. Proposal mode flags risk language and missing sections.
Profiles are the part we end up using most. A profile is a free-text description of how a specific person prefers to be communicated with — direct/diplomatic, bullets/prose, formal/casual, current sensitivities. They live in ~/.presend/profiles.json, never leave your machine, and let the checker grade against the actual recipient instead of a generic ideal.
Output is a markdown table: red/yellow/green flags, score, one-line summary. Mixed-language drafts get auto-detected and graded in their own language. Then "apply those flags" hands the corrected draft back.
Useful when: you're about to send something tricky and don't want to ask a colleague at 23:47; when you want a second pair of eyes that knows your boss's quirks; when you'd like the rewrite to happen inside your existing assistant instead of a separate web app.
Install via claude mcp add presend node /path/to/presend/dist/index.js or as a Claude Code plugin. Full setup in the repo README.