Being Human MCP server

being-human

Your assistant writes like a press release because it doesn't know who you are. This tells it.

being-human reads the prompts you've already typed, measures how you actually write, and hands the model a description of your voice specific enough to follow. It also builds a list of words to avoid, computed by contrast against the model's own output, so it's personal to you rather than borrowed from a blog post about ChatGPT tells.

Local, stdlib Python 3, no dependencies, no network calls.

my own writing, measured by this tool. every bar is one real sentence. make your own with being-human card.

quickstart

pip install being-human

being-human learn --name "Your Name"     # harvest your writing, measure it
being-human check draft.md               # AI tells
being-human verify draft.md              # is it statistically you?
being-human export --target chatgpt      # or agents, cursor, system, json
being-human card                         # draw your voiceprint as a shareable svg

No Claude Code history? Point it at a ChatGPT export or your own files:

being-human harvest --source chatgpt --path conversations.json --out .being-human/ --contrast
being-human harvest --source files --path ~/Documents/writing --out .being-human/

what comes out

Instructions with numbers attached, because "be concise" is not actionable and "your median message is 12 words" is:

- Write "i" lowercase. They do, 83% of the time.
- Vary sentence length hard. Theirs average 9 words but deviate 8, up to 81.
- No exclamation marks. Rate is 0.28 per 1k.
- Don't hedge. Hedge rate is 1.1 per 1k -- they assert.

the contrast trick

Your transcripts contain both sides of every conversation. So every word can be scored by log-odds: how much more likely you are to use it than the model is, on the same topics, in the same threads. The top of that list is your voice. The bottom is the model's.

Mine came back actually, clean, rather, genuinely, exactly, roughly. None of those appear on any published list of AI words. They're the specific tics this model has when talking to me. Yours will differ, because the contrast is against your writing.

two checks, not one

They catch different failures, and passing one says nothing about the other:

Measured on my own corpus:

Row two is the catch: slop has unremarkable grammar, so authorship distance passes it. Row three is clean prose with no tells that is plainly a different hand. The second case is the one nobody else tests for, and it's the one that bites.

Held-out chunks of real writing sit at the 49th to 60th percentile, with a 1-6% false-positive rate above p97. Thresholds are personal. Em dashes get flagged against your measured rate rather than a universal rule, which is why this doesn't punish people who genuinely use them.

mcp server

claude mcp add being-human -- being-human-mcp

Seven tools, two resources, two prompts. Configs for Claude Desktop, Cursor, Codex and Zed are in mcp/README.md.

Shortened. The full README is on GitHub.