Zero Slop

What it does

Zero Slop's MCP server is a hosted endpoint with a single tool, `deslop`. You pass it a draft (plus an optional genre and intended reader) and it returns a rewritten version with the stock phrasing removed, a writing score for the text before and after, whether names, numbers, quotes and links survived, and a review status. It suits announcements, emails, posts and reports drafted with an AI assistant that read as filler.

The work happens on the makers' Cloudflare setup, not on your machine. A TypeScript gateway takes the request, a private Python worker runs the same scorer as the open-source skill (294 weighted patterns and a 96-term lexicon), and one call goes to a hosted model, Cloudflare Workers AI by default with a single OpenRouter fallback to free models. The scorer checks the result again afterwards. If the model's edit drops or invents a fact, or the model call fails, a conservative local pass removes known stock wording instead and the result is flagged with warnings. A draft that already scores clean is sent back without a model call.

The gateway code at v2.12.25 (mcp/gateway/src/contract.ts) defaults `genre` to `general`, accepts `social`, `email`, `research` and `professional`, and caps `audience` at 200 characters. budget.ts there sets the shared model budget: 8,000 Workers AI neurons per UTC day across all users, and by default five model calls per client network per day and two per minute.

The same project also ships an agent skill, a CLI on npm (`zero-slop`) and a REST API. The npm package scores files offline and calls the hosted editor; it does not run the MCP server.

What it needs

An MCP client that can add a remote Streamable HTTP server: Claude Code, Codex, ChatGPT desktop, Claude or Cowork custom connectors, or any other remote client. No account, key or sign-in. On Claude Team or Enterprise plans an owner has to add the connector in the organisation's connector settings first.

Limits

One tool, which takes text only: it cannot read files or URLs. Drafts are limited to 20,000 characters and `audience` to 200. Model calls come from a shared free budget (8,000 neurons per UTC day overall, five calls per client network per day, two per minute, per budget.ts at v2.12.25), so it refuses calls once that budget runs out, and new calls pause in the last minute of each UTC day. Requests have a 36-second end-to-end budget. The makers say the score flags patterns and cannot tell who wrote a text. The scorer runs on Cloudflare's Python Workers runtime, which Cloudflare labels beta.

Who it suits

Suits someone who wants an assistant to tidy AI-sounding prose without installing anything, since it needs no key and runs nothing locally. It only edits text you send, so it reaches nothing else. Since every draft goes to the makers' service and on to a third-party model provider, anyone handling confidential text, or who needs more than a few edits a day, may prefer the open-source Zero Slop skill or the `npx zero-slop score` command, which score drafts offline (that needs Node.js 22+ and Python 3). If you want a writing pass without a hosted model call, the score-only route is the one to use.