No Slop

Local-first AI writing cleanup for humans and agents. No Slop detects common LLM-writing patterns, rewrites safe mechanical issues, scores the input, and flags structural problems that need human judgment.

The same deterministic text engine powers a browser UI, seven local API route families, a command-line tool, and a native Model Context Protocol (MCP) server. A Django REST Framework port provides the same core response contract for a Python backend.

Why it is useful

• Deterministic cleanup: remove em-dash overload, decorative emoji, filler openers, throat-clearing transitions, hype phrases, and inflated vocabulary.
• AI-writing detection: produce a 0-100 score with a signal breakdown.
• Human-in-the-loop review: flag risky structural patterns instead of pretending every rewrite can be safely automated.
• Agent integration: expose deslop and slop_score as native MCP tools.
• Multiple interfaces, one engine: browser, REST-style API, CLI, and MCP all call src/deslop.js.
• Local by default: the server binds to 127.0.0.1; the deterministic engine does not require a hosted service or API key.

Architecture

Browser UI ─┐
Local API ──┼──> src/deslop.js <── CLI
MCP server ─┘          │
                      └── deterministic cleanup + score + flags

Django/DRF API ──> Python port with the same response contract

Quick start

Requires Node.js 20 or newer.

git clone https://github.com/eatusc/no-slop.git
cd no-slop
npm install
npm run dev

Open http://localhost:4242.

HTTP API

The main integration endpoint accepts raw text or JSON:

curl -s -X POST http://localhost:4242/api/deslop \
  -H "Content-Type: application/json" \
  -d '{"text":"Furthermore, our seamless platform leverages cutting-edge AI. 🚀"}'

Example response:

{
  "clean": "Our smooth platform uses advanced AI.",
  "slop": {
    "score": 100,
    "label": "Pure slop",
    "signals": {
      "transitions": 1,
      "emoji": 1,
      "words": 3
    }
  },
  "fixes": {
    "total": 5,
    "byCategory": [
      { "label": "Emoji", "count": 1 },
      { "label": "Throat-clearing transitions", "count": 1 },
      { "label": "Inflated words", "count": 3 }
    ]
  },
  "flags": [],
  "words": {
    "in": 8,
    "out": 6
  }
}

Return only clean text for shell pipelines:

curl -s -X POST "http://localhost:4242/api/deslop?clean=1" \
  --data-binary @draft.md > draft.clean.md

The application also has local routes for AI-assisted rewrites, learned style examples, rule consolidation, the example corpus, and editable documentation. See the complete route and security reference in API.md.

Native MCP server

No HTTP server is required for MCP. The stdio server imports the core engine directly and exposes two tools:

Register it with Claude Code:

claude mcp add -s user noslop -- node /ABSOLUTE/PATH/no-slop/mcp/server.mjs

Claude Desktop or Cursor configuration:

{
  "mcpServers": {
    "noslop": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/no-slop/mcp/server.mjs"]
    }
  }
}

Test the full MCP client/server exchange:

npm run test:mcp

See mcp/README.md for details.

CLI

The CLI works without a server:

echo "We leverage a robust solution." | npm run deslop -- --report
node cli/deslop.mjs draft.md > draft.clean.md
node cli/deslop.mjs --json < draft.md

Example macOS clipboard workflow:

pbpaste | node cli/deslop.mjs | pbcopy

Local AI rewrite and voice learning

The deterministic engine handles changes that are safe to automate. For heavier rewrites, POST /api/rewrite orchestrates a locally installed Claude or Codex CLI. How the pipeline works:

• Your CLI login, no API keys. The server spawns claude -p or codex exec as subprocesses and rides whatever authentication those tools already have. It never accepts, stores, or forwards a hosted API key. Codex runs in a read-only sandbox with an ephemeral session; both engines get a hard 180-second timeout with stdin closed.
• Few-shot retrieval without dependencies. Every time you accept a rewrite you like (or hand-edit one), the before/after pair is saved locally via POST /api/style. On the next rewrite, the server tokenizes the input, builds term-frequency vectors, and ranks all learned pairs by lexical cosine similarity (with a stopword list, no embeddings, no libraries), then packs the seed examples plus the top 8 most relevant learned pairs into the system prompt. The style bank scales to hundreds of examples without bloating the prompt.
• Consolidate voice. Once you have at least 3 saved edits, POST /api/consolidate sends the whole before/after corpus back through the Claude CLI and asks it to distill 8-15 concrete imperative rules from your actual editing patterns. The result is written into a marker-bounded section of voice.md (<!-- LEARNED:START --> to <!-- LEARNED:END -->) with an atomic tmp-file rename, so repeated consolidations replace only that section and the hand-written voice guide around it is never touched.
• Prompt hygiene. The system prompt sends a distilled "tells to avoid" list instead of the full word-swap table (which pushes models toward literal substitution instead of rewriting), and strips the quoted example sentences from voice.md so the model cannot parrot them as openers.

The associated route families are:

• POST /api/rewrite
• GET, POST, and DELETE /api/style
• POST /api/consolidate
• GET /api/examples and POST /api/examples/dismiss
• GET and POST /api/doc/{rules|voice|api}

Personal learned examples are gitignored. AI rewrite routes require a locally installed and authenticated CLI; the deterministic engine, API, CLI, and MCP tools do not.

Django REST Framework implementation

django_api/ ports the engine and API contract to Python, Django, and Django REST Framework. It includes:

• JSON and text/plain request parsing
• request validation and a 4 MB limit
• local-development CORS controls
• anonymous rate limiting
• 18 endpoint, regression, throttling, and parity tests

cd django_api
python3.12 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python manage.py test deslop
python manage.py runserver 127.0.0.1:8420

Parity between the JavaScript and Python engines is enforced in CI: scripts/parity-check.mjs runs both engines over shared fixtures plus the full example corpus and fails on any output difference. See django_api/README.md for implementation notes.

Development and validation

npm run build       # production Vite build
npm test            # Node engine regression tests
npm run test:mcp    # real MCP stdio client/server smoke test
npm run test:parity # diff Node and Python engine output on shared fixtures
npm run check       # all Node checks

CI runs the Node build, engine tests, MCP smoke test, cross-engine parity check, and Django test suite on every push and pull request to main.

Security model

The local server can invoke installed AI CLIs and update local style/doc files, so it intentionally binds only to loopback. API requests with a non-localhost browser Origin are rejected. Do not reverse-proxy or expose port 4242 to a network without adding authentication and authorization.

See SECURITY.md for reporting and deployment guidance.

License

Code is available under the MIT License. Example captions are sample content. Third-party essay text is not included; the long-form writing rules are general writing guidance.