Zero Slop
github.com/manavmishra/zeroslop- Category
- AI Agents
- Rank
- No. 1229Tools index
Previous survey · No. 1198 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Productivity & Collaboration
- Interfaces
- Agent Skill / Plugin
- GitHub
- 168 stars
- Latest release
- v2.12.40
- Date
About
Zero Slop is an open-source agent skill that finds and removes AI-generated filler from writing. The repository includes packaging for Claude and Codex plugins, an MCP configuration, a pre-commit hook and a GitHub Action.
What it does
Zero Slop is an editing kit for prose that sounds machine-written. A local Python scorer rates a draft from 0 to 100 using a database of weighted regular expressions and a word list, so a lower number means fewer stock phrases. Your existing AI assistant does the actual rewriting, following an eight-step editorial routine, while local checks confirm that names, numbers, quotations and links survived the edit. The same scorer runs from a command line, in a GitHub Action that can fail a build, and behind a free hosted endpoint for clients without skill support.
Why it's ranked here
This is a writing skill with real tooling behind it. The MIT-licensed repository carries an npm command line tool, a dependency-free Python wheel, a composite GitHub Action that makes no network request, and a hosted server reachable through MCP (the Model Context Protocol) or plain REST. The documentation is unusually candid: it says the score measures phrasing patterns rather than authorship or truth, and the benchmark notes admit the drafts were synthetic, the judges were language models, and inter-rater agreement was weak. That honesty about limits, paired with broad install surfaces, is the case for it.
What's good
The fact gate is the useful idea. A rewrite that drops or invents a name, number, quote or link is rejected by a local check instead of trusted to the model. The installer refuses to write into your home or root directory, stages the new copy beside the old one, and restores a backup if the final swap fails. The command line editor never overwrites the source file and prints results to standard output. The REST docs spell out six result statuses and give an exact approval condition, so an automated pipeline can stop on a warning rather than shipping a flagged edit.
Tradeoffs
The evidence for quality is mostly self-produced. The benchmark notes state that no human raters took part, the corpus had 50 synthetic drafts, and the two judging passes agreed on winners only about half the time, with the judge prompt not preserved. The hosted editor is tight: five model calls per client network per day, two per minute, shared across every channel. Offline scoring needs Python 3, hosted editing needs Node 22, and Linux Homebrew is untested. The skill's own rules concede the scorer sees only the lexically anchored subset of tells, so a clean number is not a clean draft.
How to use it well
Its best fit is a writing team that already drafts with Claude Code, Codex or Cursor and wants a repeatable check. Install the skill, run its inspect mode first to see flagged passages, then let your assistant edit. Put the GitHub Action on documentation folders with the default threshold of 25, or set it to report without failing while you calibrate. Treat the hosted endpoint as a fallback, given its daily cap. Do not use it to decide who wrote a text: the project says plainly that the score cannot identify an author.
Technical notes+
bin/zero-slop.mjs is the npm entry point: it copies SKILL.md plus the references, scripts and data folders into a per-harness skills directory (claude, codex, cursor, opencode, zed), rejects the filesystem root, home and working directory as targets, and only overwrites an existing install whose frontmatter names zero-slop and which holds the scorer and data/patterns.json. Its score subcommand spawns python3 on the bundled scorer; deslop caps input at 128 KiB, requires valid UTF-8, and bounds the timeout to 300 seconds. data/patterns.json stores each rule as a name, category, case-insensitive Python regex and weight. action.yml passes inputs through env rather than expression interpolation and treats a folder with no Markdown or text files as a pass. mcp/gateway/src/index.ts registers a single deslop tool with a noauth security scheme, reads request bodies under a size cap with a 10 second deadline, sets no-store and DENY frame headers, and fails its health check unless the scorer version matches and the editor secret is at least 32 characters. pyproject.toml declares an empty dependency list.
Observed
- License
- MIT
- Languages
- Python scorer (standard library only), JavaScript command line tool, TypeScript hosted gateway
- Packaging
- npm package with a CLI binary; PyPI wheel with no dependencies; Homebrew tap for macOS
- Interfaces
- Agent skill, CLI, GitHub Action, hosted MCP server, REST API
- Runtime requirements
- Python 3.9 or newer for offline scoring; Node.js 22 or newer for hosted editing
- Hosted access
- Free, no API key, no uptime SLA
- Benchmark method
- LLM judges on synthetic drafts; no human raters, per the repository's own benchmark notes
Read from README.md, package.json, pyproject.toml, LICENSE, docs/cli.md, docs/rest-api.md, bin/zero-slop.mjs, data/patterns.json, mcp/gateway/src/index.ts, SKILL.md, action.yml, bench/README.md.
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.