MCP Serverstreamable HTTP

Use Rephrasy where you already write — inside your AI assistant.

Add one URL as a connector and your assistant can rewrite drafts and score them for AI-written content mid-conversation. No SDK, no glue code, no second tab. Every call runs on your Rephrasy account and the credits you already have.

Connector URL

https://v2-humanizer.rephrasy.ai/mcp

Authenticate with the same Rephrasy API key you use for the REST API.

Works withClaudeChatGPTCursorClaude Code

What you can do with it

Four things people actually use it for, with the responses the tools really return.

You say

Humanize this blog intro for me, it reads like a press release.

Your assistant callshumanize_text

You get back

{
  "output": "Remote work didn't arrive gently. It landed in March 2020…",
  "engine": "v4",
  "report": {
    "words_in": 118,
    "words_out": 121,
    "length_ratio": 1.025,
    "figures_kept": ["2020", "42"],
    "figures_missing": [],
    "figures_added": [],
    "flesch_reading_ease": 68.4
  }
}

Every rewrite comes back with a preservation report: how the length moved, which figures survived, and which the rewrite lost or invented. You see what changed instead of trusting that nothing did.

And here it is running

A real conversation, not a mock-up: ChatGPT calls the connector over OAuth, rewrites the sentence with v4 and reports what survived — the figure 2024 kept, nothing lost or invented.

ChatGPT using the Rephrasy connector to rewrite a sentence

Set it up

One URL and your API key. Under a minute in any client.

Open Claude and go to Settings → Connectors → Add custom connector.

Click a step to see that screen.

Your API key

Authorization: Bearer YOUR_API_KEY

The same key as the REST API — find it in your Rephrasy account. Claude and ChatGPT ask for it once on our consent page and get a token back, so the key stays with us. Claude Code and Cursor send it directly as the header above. Either way it is never part of a prompt.

Tools and pricing

Five tools. All need your Rephrasy account; two spend a credit, three do not.

ToolCostWhat it does
humanize_text1 creditRewrites a draft into natural prose. Takes a purpose (blog, marketing, seo, report, technical) and picks the engine for you, or name the engine yourself. Returns the rewrite plus a preservation report.
check_ai_score1 creditScores text for AI-written content and returns the passages that scored highest, so you can see which parts to revise.
recommend_engineNo creditSuggests the engine and voice for a kind of writing.
readability_scoreNo creditFlesch reading-ease score for any text.
list_enginesNo creditLists the engines, voices and detectors your connector can use.

Models

The same model names the REST API takes. Leave the choice to the purpose you give, or name one directly.

v4

v4 — the recommended default. Natural rhythm and sentence variety, length stays close to the source.

ft_clean-v3

ft_clean-v3 — the highest meaning retention (0.96 mean cosine on our benchmark against 0.92 for v4) and near-identical length. For technical writing and anything with figures.

v3

v3 — a voice-driven rewrite. Pick a style; 'creative' is the strongest and also the longest, so avoid it when length matters.

seo

seo — rewrites around the keywords you need to keep.

Detectors

The Rephrasy AI Detector works the moment you connect — no extra account, one flat credit per check. You can also run any other detector on your own key: add it to the connector as a request header and that provider bills you directly, with no Rephrasy credit charged.

DetectorYour key
Rephrasy AI DetectorIncluded — nothing to set up
wasitaigenerated.comx-wia-key
GPTZerox-gptzero-key
Winston AIx-winston-key
Copyleaksx-copyleaks-key + x-copyleaks-email
Pangramx-pangram-key
ZeroGPTx-zerogpt-key

Ask list_engines to see which detectors your connector can currently use. Pangram needs at least 100 words; Copyleaks at least 255 characters.