Typographer, inside your agent.
The same deterministic linter and humanizer that power Typographer, exposed as Model Context Protocol tools. Your AI agent can see the machine tells in its own draft and rewrite until the prose reads human, in a loop, with no person in the middle.
What it is
MCP is the open standard that lets an AI assistant call outside tools while it works. The 701am server gives three tools to any MCP host, from Claude Desktop to Cursor to Cowork, turning our writing engine into something an agent can call mid-task.
Three tools
score_text
Returns the three pillars (voice, readability, information gain) and every tell it found. Deterministic and cheap.
find_ai_tells
Just the machine tells: the empty vocabulary, the hedging, the flat rhythm. The fastest signal to act on.
humanize
Rewrites the text to strip the tells, with an adversarial self-revision loop. This one calls a model, so it counts against your plan.
Every tool takes an optional mode, one of these five: general, marketing, technical, academic, news. It tunes the linter to what is being written.
Add it to Claude Desktop
Drop this into your claude_desktop_config.json, restart, and the three tools appear:
{
"mcpServers": {
"typographer": {
"command": "npx",
"args": ["-y", "@701am/typographer-mcp"],
"env": { "TYPOGRAPHER_API_KEY": "tg_live_xxx" }
}
}
}
It works the same in any MCP host that speaks the protocol. Point it at a different deployment with TYPOGRAPHER_API_URL.
The loop it enables
Draft with the model, score it, read the tells when the voice is low, rewrite, then score again. Because the score is a deterministic check and not another model's opinion, the agent can trust it instead of gaming it. That is the whole point: a verification primitive an agent can call, not a black box that grades on vibes.
Get a key
Sign in at 701am.com and create an API key in your account. Drop it in as TYPOGRAPHER_API_KEY. The two read tools are free and deterministic; humanize counts against your plan.