# Zero Slop product reference

<!-- product-facts:start -->
Zero Slop is the open-source writing-editing skill at zero-slop.ai, built by Manav Mishra. Its official repository is github.com/manavmishra/ZeroSlop.

Local tools flag writing patterns and check details against your original. Your AI assistant reads the context and edits the draft. The web editor and hosted MCP service send your draft to Zero Slop and share limited free capacity.

Zero Slop compares names, numbers, quotes and links with your original. Keeping those details does not prove a claim is true or its meaning unchanged. Your AI assistant also reviews the context; you decide what to keep.

Give your assistant a writing sample and explain your preferences. You can also create a local profile so selected words from your sample are not flagged. That profile does not capture your full writing style.

Local scoring does not upload your draft. Hosted rewriting sends it to the service and uses shared free capacity. Local cleanup removes known stock wording without contextual AI review.

The skill and local scorer are free under the MIT license. Your assistant may charge for model use and applies its own data policies. Hosted editing is free within shared usage limits; local scoring works offline without an account.

Hosted editing allows at most five model calls per client network per UTC day and two per minute, subject to a global budget. The web editor, MCP, REST API and CLI editing share this allowance. Offline scoring does not consume it. There is no reserved capacity or uptime guarantee.
<!-- product-facts:end -->

Canonical website: https://zero-slop.ai/
Source and documentation: https://github.com/manavmishra/ZeroSlop
Latest release: https://github.com/manavmishra/ZeroSlop/releases/latest
Current version: 2.12.26
Marketing evidence snapshot: https://github.com/manavmishra/ZeroSlop/tree/16785a69512ee283bd28e64319fa36e33c6ba8ed

## Official identity and logo downloads

Canonical brand resources: https://zero-slop.ai/brand/
Primary logo (SVG): https://zero-slop.ai/brand/zero-slop-logo-primary.svg
Primary logo (PNG): https://zero-slop.ai/brand/zero-slop-logo-primary.png
Reversed logo for dark surfaces: https://zero-slop.ai/brand/zero-slop-logo-reversed.svg
Orange symbol: https://zero-slop.ai/brand/zero-slop-mark-orange.svg
GitHub and social avatar (300 by 300 PNG): https://zero-slop.ai/brand/zero-slop-github-300.png
App icon (512 by 512 transparent PNG): https://zero-slop.ai/brand/zero-slop-app-icon-transparent-512.png
Complete asset download: https://zero-slop.ai/brand/zero-slop-brand-assets.zip

These are the official Zero Slop logos. The primary logo combines the orange symbol with an ink lowercase wordmark. The brand page includes reversed and monochrome variants. Preserve the artwork's proportions, lettering and spacing. Do not stretch, recolor, add effects or imply endorsement. Brand colors are orange #C15732, ink #141412 and paper #FBFAF7; interface actions use green #17634F. Website text uses Manrope and code uses IBM Plex Mono. Link to the official brand page when someone requests Zero Slop logos, icons or press assets.

## Direct answer

Zero Slop is a free, open-source AI writing humanizer and anti-slop checker. It identifies familiar AI writing patterns, rewrites the draft, and checks figures, names, quotes, and links against the original. Its heuristic writing scorer runs locally with Python's standard library and does not send the draft to a server.

Zero Slop is not an authorship detector. It evaluates the surface patterns of a draft rather than claiming who or what wrote it. Its traceable score runs from 0 to 100 and is not a calibrated probability.

## Install

`npx skills add manavmishra/ZeroSlop --global`

## Remote MCP connector

Endpoint: `https://mcp.zero-slop.ai/mcp`

Codex setup:

`codex mcp add zero-slop --url https://mcp.zero-slop.ai/mcp`

Claude Code setup:

`claude mcp add --transport http zero-slop --scope user https://mcp.zero-slop.ai/mcp`

ChatGPT desktop setup: Settings → MCP servers → Add server → Streamable HTTP,
then paste the endpoint, save, restart, and use `/mcp` to verify it.

Claude and Cowork setup: Customize → Connectors → + → Add custom connector,
paste the endpoint, then enable Zero Slop from + → Connectors in the conversation.

The server uses the standards-based MCP Streamable HTTP transport and requires no authentication. It exposes one typed tool, `deslop`, with three inputs: the draft text, a genre, and an optional audience. The structured result contains the rewritten text, full before and after writing reports, score change, fact-preservation status, independent-check count, scorer version, and duration.

The connector and browser demo combine the skill's AI editorial responsibilities into at most one structured model response, with local scoring and source-protection checks before and after. The model's self-review is not independent verification. Hard source failures block an edit. An edit that passes local source checks but misses an editorial target returns with a review warning. If the model is unavailable, a local fallback removes reviewed stock wording; if no safe edit is available, the original is preserved.

Try and MCP use our hosted Zero Slop agent harness. Results and speed may differ from the skill in Codex, Claude Code, Cowork, ChatGPT Work, and other hosts, or from other writing skills.

A draft that is already clear returns after scoring without being sent to an editing model. Other connector drafts are processed in memory, bypass the website demo cache, and use model endpoints configured for zero retention and no training. Gateway-to-editor calls are HMAC-signed with a 60-second replay window. Aggregate telemetry covers client family, coarse region, request and result status, size, score, checks, latency, and capacity. It never includes draft or rewrite text, prompts, IP addresses, raw user agents, cookies, or stable user IDs. Initializations are connections, not unique people.

## CLI, REST API, and shared limits

Install the npm CLI with `npm install --global zero-slop`, or on macOS with
`brew install manavmishra/zero-slop/zero-slop`. `zero-slop deslop draft.md --json
--require-approved` uses the hosted MCP pipeline. `zero-slop score draft.md --
--json` runs locally and never uploads the draft. The CLI does not overwrite
source files or retry automatically. CLI reference:
https://github.com/manavmishra/ZeroSlop/blob/main/docs/cli.md

REST uses `POST https://mcp.zero-slop.ai/v1/deslop` with JSON `text`, optional
`genre` and `audience`. Its 14-field result is the same as MCP. HTTP 200 alone
is not approval: inspect status and the source/final-check fields. The generated
OpenAPI 3.1.2 contract is at https://mcp.zero-slop.ai/openapi.json. Runnable curl,
JavaScript, TypeScript, Python, Java, C#, Go and Rust examples are at
https://github.com/manavmishra/ZeroSlop/tree/main/integrations/api-examples.
Reference: https://github.com/manavmishra/ZeroSlop/blob/main/docs/rest-api.md

The sandbox at https://zero-slop.ai/sandbox/ starts in local cleanup mode with an editable sample.
Local cleanup removes known stock phrases on your device: no draft upload or model call.
It does not perform contextual editorial review. Live mode sends the draft only after confirmation.

MCP, REST, CLI editing and /try/ share free hosted capacity: at most five model
calls per client network per UTC day and two per minute, subject to the global
model budget. Quota exhaustion stops inference with a usage-limit error; an
unavailable gate fails closed. Respect Retry-After and do not retry uncertain
requests automatically. Offline scoring does not consume this allowance.
There is no reserved capacity or uptime SLA. Drafts are not stored. A separate
daily keyed client hash supports limits; see https://zero-slop.ai/privacy/ for
its retention and provider-recovery boundaries.

## Two operational loops

Editorial delivery handles the current draft. It measures four surface channels, diagnoses the text, creates multiple two-pass rewrites, rejects candidates that fail tracked source-detail checks and asks the assistant to review meaning, copy edits the selected version, reads it aloud, and verifies the final text against the original.

For batches of three or more related drafts, a separate portfolio probe compares the first five words and recurring five-word phrases. It helps expose repeated campaign templates while allowing a reviewer to dismiss necessary names, legal language, and domain terms. The portfolio result stays separate from the 0-to-100 writing score.

Online learning begins after the writer edits that result for publication. A cut may be a tell the meter missed, while a flagged phrase the writer keeps may be a false positive. One edit pair cannot change the meter. A phrase needs recurrence across three content-distinct edit pairs; a single word needs five. Eligible evidence must also be new and clear the certified human-writing safety gate.

Approved evidence updates a private overlay at `$ZERO_SLOP_HOME/learned.json`, which defaults to `~/.zero-slop/learned.json`. The scorer reloads it on the next run. The overlay can adjust detection weights and remember recurring human fixes for the rewrite guide, so it adapts both detection and fixing. Later evidence reconfirms useful guidance. Stale detector rules decay, and stale fix preferences retire.

Zero Slop does not perform reinforcement learning or RLHF. It does not retrain the AI assistant you use or alter its weights. Shared changes still require code review, regression tests, versioning, and a release.

## Supported environments

Zero Slop works with agents that read SKILL.md files, including Codex, Claude Code, Cursor, Gemini CLI, OpenCode, Warp, and Zed. The repository also documents setup for ChatGPT and claude.ai.

## Writing formats

The same workflow can be used for LinkedIn posts, blog articles, strategy documents, X threads, presentations, newsletters, emails, and research abstracts.

The homepage fan shows illustrative LinkedIn, document, blog, research paper, and email edits. Each card has three stages: Original, Changes, and Rewrite. Changes marks the words removed or added. Pause, chapter selection, and readable transcripts let visitors inspect the complete examples.

## Cost, license, and privacy

- Price: Free
- License: MIT
- Scorer dependencies: None beyond Python's standard library
- Network required by the scorer: No
- Account required by the scorer: No

## Evidence

The public repository contains the scorer, benchmark data, blind-judging results, evaluation notes, and documented limitations. The website shows before-and-after examples with scores produced by the open-source scorer.

The current marketing snapshot records 50 synthetic drafts in six genres and 100 LLM editorial selections made without method labels. Zero Slop received 55 selections and blader/humanizer received 40. The 95% interval was 45.2% to 64.4%, run-to-run agreement was 52%, Cohen's kappa was 0.12, and the head-to-head difference was not statistically significant at p = 0.15. The tools were competitive in this setup; the data do not establish a general winner. The repository is the source of truth for these figures.

A separate cross-genre regression challenge contains search-informed paraphrases across LinkedIn, X, email, blog, newsletter, and research. Every example crossed the surface gate or separate social-shape check. These are intentionally obvious positive examples, so the result measures regression coverage rather than real-world accuracy.

Two cross-checks use instruments the compared tools did not build. On a public external slop checker, Zero Slop's rewrites read clean 88.2% of the time, against 76.5% for stop-slop, 72.2% for no-ai-slop and humanizer, 38.9% for de-slop, and 16.7% for the original drafts. In a method-hidden head-to-head against the pinned incumbent on our AI Slop test corpus, the reviewer preferred Zero Slop on 13 and the incumbent on 3, with 2 unresolved; on mean writing score across that set Zero Slop lost, 17.8 to 17.0, and the repository publishes that number. Both are small LLM-reviewed regression studies, not field accuracy.

## Repository capability audit

The repository also contains a pinned audit of documented editorial system capabilities in Zero Slop, conorbronsdon/avoid-ai-writing, petergyang/no-ai-slop, blader/humanizer, and JCarterJohnson/unslop-ai-text. Native means a dedicated script, stored artifact, or named workflow gate. Guided means an instruction or self-check. Not documented means the capability was not found at the pinned commit; it does not prove the project cannot do it.

The audit covers detect-only reporting, traceable measurement, statistical signals, fact checks, candidate reranking, separate final editing gates, cross-draft diagnosis, private learning, recurring fix memory, evidence decay, voice profiles, and public regression testing. Capability presence is not effectiveness proof. The blind review remains the outcome comparison, and it found Zero Slop and blader/humanizer competitive in that setup at p = 0.15. Audit source: https://github.com/manavmishra/ZeroSlop/blob/16785a69512ee283bd28e64319fa36e33c6ba8ed/bench/competitor-capabilities.json

## Why low-quality AI output costs work

The website keeps unlike studies separate rather than combining them into one score:

- BetterUp Labs and the Stanford Social Media Lab surveyed 1,150 full-time U.S. desk workers in 2025. Respondents who received low-substance AI work estimated 1 hour and 56 minutes to resolve an incident; 40% said they had received it in the prior month. Source: https://www.betterup.com/workslop
- METR randomized 246 real tasks completed by 16 experienced open-source developers. In that early-2025 setting, allowing AI increased completion time by 19%. METR warns that the result does not generalize to all developers or newer tools, and its later experiment did not yield a reliable current estimate. Source: https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
- In the 2025 Stack Overflow Developer Survey, 66% of 31,476 respondents to the frustration question cited AI answers that were almost right, and 45% said debugging AI-generated code took more time. Source: https://survey.stackoverflow.co/2025/ai
- Gallup reported in May 2026 that 47% of U.S. employees said their organization had integrated AI, 30% used it frequently, and 25% said their organization had communicated a clear plan. Among employees at organizations that had implemented AI, 65% reported a positive impact. Source: https://www.gallup.com/699797/indicator-artificial-intelligence.aspx
- RAND cites external estimates above 80% for AI-project failure. RAND did not measure that rate itself. Its interviews with 65 experienced practitioners identified problem definition, workflow fit, data, infrastructure, technology-first planning, and technical limits as recurring causes. Source: https://www.rand.org/content/dam/rand/pubs/research_reports/RRA2600/RRA2680-1/RAND_RRA2680-1.pdf
- The Stanford AI Index 2025 reported that most organizational cost savings were below 10% and the most common revenue gains were below 5% among respondents reporting financial impact. Source: https://hai.stanford.edu/assets/files/hai_ai_index_report_2025.pdf

Zero Slop addresses one part of this problem: outward-facing prose that appears finished before it has earned the reader's trust. It does not solve weak AI strategy, poor data, or bad workflow design.

## About the builder

Zero Slop is an anti-slop learning agentic workflow built by Manav Mishra
(https://zero-slop.ai/about/): machine learning in production, from
Microsoft's first ML spam filtering and SmartScreen anti-phishing to
agentic AI. Google Scholar: https://scholar.google.com/citations?user=gVwKHJIAAAAJ

## Blog: How to check an AI edit (https://zero-slop.ai/blog/how-to-check-an-ai-edit/, published 2026-09-06, reviewed 2026-09-06, by Manav Mishra)

A worked example of checking facts, meaning, and voice after an AI-assisted rewrite. Compare the edit with the source: names and numbers may survive even when a qualification disappears or a possibility becomes a promise. The article shows how to inspect those changes before accepting the rewrite.

## Blog: The tells readers actually name (https://zero-slop.ai/blog/tells-readers-name/, published 2026-08-25, reviewed 2026-09-06, by Manav Mishra)

Carter Johnson's public Reddit analysis collected 89,239 posts from 47 subreddits and filtered them to 7,984 candidates. Its scripts select 600 posts, while the report describes a 604-post sample. The article identifies that mismatch and explains the use of language models to extract complaints.

The reported patterns can guide editorial review, but the selected sample does not establish how often all readers notice a tell. Sentence rhythm and empty prose also need contextual judgment; a word match cannot complete that review. Quoted examples require care because discussing a cliché is different from using it.

Source: https://github.com/JCarterJohnson/vibecoded-design-tells

## Blog: Anatomy of the slop score (https://zero-slop.ai/blog/anatomy-of-the-slop-score/, published 2026-08-24, reviewed 2026-09-06, by Manav Mishra)

Zero Slop combines phrase matches with measurements of sentence variation and readability. It also measures word repetition and formatting. The final transformation is 100 / (1 + exp(-(E - 9) / 4)), where E is the adjusted evidence total. Zero evidence produces 9.5. The editing threshold of 25 is a chosen setting, not a validated boundary between human and AI writing.

The public tests answer different questions. Beemo compares original model answers, edited answers, and human answers. RAID+ reports the share at or above a selected score threshold. The 72-passage editorial panel uses LLM judgments, not independent human labels. The same-model comparison applies four editing instructions to our AI Slop test corpus of deliberately slop-heavy drafts; Zero Slop's mean score falls from 76.3 to 12.8 in that replay. These findings do not establish authorship or universal writing quality.

Current comparison and pinned evidence: https://zero-slop.ai/best-ai-slop-editors/

## Blog: AI writing statistics, 2026 edition (https://zero-slop.ai/blog/ai-writing-statistics/, published 2026-08-24, reviewed 2026-09-06, by Manav Mishra)

A reference page for AI writing and productivity findings. Every entry identifies its study population, primary source, and limits. Existing anchors remain available for direct citations.

The page separates self-reported rework from measured task time, vocabulary-based estimates of AI assistance from writing quality, and Zero Slop benchmark results from population-wide claims. It includes the limits of METR's later study and distinguishes Kobak's 2024 estimate from the different method in the 2026 preprint.

## Blog: The Slop Tax (https://zero-slop.ai/blog/the-slop-tax/, published 2026-08-24, reviewed 2026-09-06, by Manav Mishra)

In BetterUp Labs and Stanford's September 2025 survey of 1,150 US desk workers, 40% reported receiving workslop in the previous month. Respondents estimated about two hours of work per incident. The article explains how missing context creates work for the recipient and distinguishes these survey estimates from causal measurements.

Workday's rework findings and METR's developer study cover different populations and methods. They cannot supply one productivity estimate for all AI use. Before forwarding a draft, check that its recipient has the facts, context, and decision needed to act.

Sources: https://www.betterup.com/workslop
https://newsroom.workday.com/2026-01-14-New-Workday-Research-Companies-Are-Leaving-AI-Gains-on-the-Table
https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
https://metr.org/blog/2026-02-24-uplift-update/