ToHuman for ChatGPT: the humanizer plugin, skill, and MCP connector

Custom MCP connector — advanced, and you bring your own key

OpenAI's Apps SDK documents connecting an MCP server to ChatGPT in developer mode. Point that flow at https://tohuman.io/mcp and send the header Authorization: Bearer YOUR_KEY, using a key from your ToHuman account. Two things to know first. OpenAI gates developer-mode connectors by account and workspace policy, so availability is theirs to decide, not ours — and the client we document end to end is Claude, not ChatGPT. And the connector needs your own API key: the server answers a tools/list without authentication, but a humanize call without a key returns This tool needs a ToHuman API key (checked again September 7, 2026). There is no paste-the-URL-and-go mode. Keys are free; the API quickstart has the 30-second version.

The humanize-text skill for Codex, Claude Code, Cursor and Copilot

Since August 11, 2026 we publish humanize-text, an Agent Skill in the open SKILL.md format — developed by Anthropic, released as an open standard, and now read by Codex, Claude Code, Cursor and GitHub Copilot. If you want the writing side rather than the tooling side, how to humanize ChatGPT text covers the manual and prompt-level methods.

What is a humanizer skill?

A skill is a folder, not an app: instructions the agent loads only when a task matches — "the user asked me to humanize something" — plus a stdlib-only Python script that POSTs the text to our API. Ours bundles the intensity guidance, the chunking rule and an error table, so the agent knows what to do on a 429 instead of guessing.

How do I add the humanize-text skill?

curl -L -o /tmp/ht.zip https://tohuman.io/skills/humanize-text.zip
unzip /tmp/ht.zip -d ~/.claude/skills/   # Codex reads ~/.agents/skills/
export TOHUMAN_API_KEY=YOUR_KEY          # free plan: 2,500 words a month

Then ask your agent to humanize something and it loads the skill itself. The limits, plainly: it runs on your own key, so the free plan's 2,500 words a month is the ceiling. It caps at 2,000 words per call. It is slow — our smoke tests on one 69-word paragraph took 8 seconds on medium and 118 seconds on minimal, so budget minutes, not seconds. And it is not listed in any skill directory yet: the zip is the install path, which means nobody is vetting it for you but us.

MCP for Claude, Cursor and other clients

Outside ChatGPT, the connector is one command. The endpoint is a hosted remote server — nothing to run locally.

claude mcp add --transport http tohuman https://tohuman.io/mcp \
  --header "Authorization: Bearer YOUR_KEY"

It exposes one tool, humanize(text, intensity) — intensity being minimal, subtle, medium (default) or heavy — with a ceiling of 2,000 words per call. The same config shape works in Claude Desktop, Cursor, Windsurf and Codex; clients that only launch local stdio servers can bridge through the open-source mcp-remote adapter. For per-client config files, use the per-client MCP setup tutorial; for the reasoning, read why we host an MCP server at all.

Humanizing code ChatGPT wrote

Is there a tool that can convert ChatGPT code to human code?

Not this one, and it is worth being precise about why. ToHuman is a prose humanizer — the model is fine-tuned on sentences, and what it changes is sentence rhythm, word choice and structure. What makes generated code look generated is different: boilerplate, naming conventions, over-explained comments, defensive scaffolding nobody asked for. Renaming variables is refactoring, not humanizing, and we will not claim a rewrite of your logic that we cannot do.

What it does do is the prose that ships with the code, which is usually where the AI tell actually lives: the README, the docstrings, the commit message, the PR description, the changelog entry, the comments — the parts a reviewer reads as writing. Those are complete sentences, the input shape the model was trained on. Paste a function body in and you get a mangled function body back.

The practical version at volume: humanize the documentation step in your pipeline, not the source file. The deeper pipeline route shows how to wire the call into a script instead of a UI.

ChatGPT humaniser (UK spelling)

If you searched for a ChatGPT humaniser with an "s", you are in the right place — same tool, we just spell it the American way in our own copy. British, Irish and Australian English are handled the same as US English: the humanizer does not force-convert your spellings, so "organised" stays "organised" unless you ask for a heavier rewrite that changes the wording anyway.

What it costs

Every route on this page draws on the same account allowance — plugin, skill and connector spend from one balance, so there is nothing to buy twice.

• Free — 2,500 humanized words a month, no credit card. Includes API access, which is what all three routes use.
• Pro — $19/month for 100,000 words.
• Word pack — $9 one-time for 50,000 words that never expire, on top of your monthly allowance.

Past the allowance the API returns HTTP 402 with your usage and an upgrade link, rather than failing silently. Detail on the pricing page.

FAQ

Does ChatGPT have a humanizer tool?

Sort of. OpenAI hosts an "AI Humanizer" page, and ChatGPT will rewrite its own output if you ask. The catch is the same model grading its own homework — the second draft comes out of the same distribution as the first, which is why re-prompted text reads better while still sounding like it came from the same place. A dedicated humanizer is fine-tuned only for the rewrite. Both help, differently.

What is the best humanizer for ChatGPT?

It depends where you work. In the chatgpt.com window, you want a browser plugin that rewrites a selection in place. In an editor or an agent, you want a skill or an MCP connector your assistant can call mid-task. Cross-vendor quality claims are hard to trust — everyone benchmarks against detectors that get retrained without notice. Run your own paragraph through two tools and read them.

Is ToHuman listed in the ChatGPT plugin directory?

Not yet. We have built and packaged an MCP-based app for OpenAI's directory and the endpoint is live at https://tohuman.io/mcp, there is no listing yet: the submission itself is queued behind our own OpenAI organisation verification. Until it is approved, the working routes are the browser plugin, the humanize-text skill, and a connector you add with your own API key. We will not call it listed before it is.

Is there a humanizer skill for Claude Code?

Yes — humanize-text, in the open SKILL.md format that Claude Code, Codex, Cursor and Copilot all read. Unzip it into ~/.claude/skills/ (~/.agents/skills/ for Codex), set TOHUMAN_API_KEY, and the agent loads it when a task matches. It caps at 2,000 words per call, and no skill directory has vetted it.

Is there a tool that can convert ChatGPT code to human code?

Not ToHuman, and we would rather say so. It is a prose humanizer: it changes sentence rhythm, word choice and structure. What makes generated code look generated is boilerplate, naming and comment style, and rewriting those is refactoring. Where it helps is the writing that ships with the code — READMEs, docstrings, commit messages, comments.

Can AI detect a humanizer?

Detectors do not detect tools. They score statistical properties — perplexity, burstiness, token-level predictability — and return a probability, not a verdict. Those scores are noisier than the marketing on either side suggests: in our own 861-sentence GPTZero study, 13.8% of verified human-written sentences were flagged AI or mixed, rising to 19.8% on news prose. No tool can promise a score from a closed classifier, and we publish no bypass guarantees.

The numbers above come from our own 861-sentence GPTZero study.

Start with the paragraph that is bothering you

The fastest test takes thirty seconds: take the ChatGPT output you already do not like, run it through, and read both versions side by side. A free API key gets you the plugin, the skill and the connector — 2,500 humanized words a month, no credit card — or paste 700 characters on the homepage with no account at all.