Claw-ED
Your AI co-teacher that writes like you teach. One command, 9 files.
Claw-ED is a command-line tool that helps teachers generate lesson plans, assessments, slides, interactive games, and more from their own curriculum files. It learns your teaching style from your existing materials and produces content that sounds like you wrote it.
It works with multiple AI providers (Anthropic, OpenAI, Google, Ollama) so you can pick what fits your budget. Your files stay on your machine. The only external call is to the AI provider you choose.
Requires Python 3.10+. Works on macOS, Windows, and Linux.
Install
$ pip install clawed
# run the setup wizard (picks your provider, creates workspace)
$ clawed setup
# ingest your curriculum files
$ clawed ingest ./my-lessons/
# generate a lesson
$ clawed lesson "The Civil War"
The setup wizard is interactive. It walks you through provider selection, API key configuration, and workspace creation. If you have curriculum files ready, it'll ingest those too.
What it looks like
$ clawed lesson "Causes of World War II" --grade 10 --standards Generating lesson package... Standards: CCSS.RH.9-10.2, CCSS.RH.9-10.6 Quality gate: 12/12 checks passed Saved 9 files: teacher_plan.docx (lesson plan + answer keys) student_packet.docx (handout, no answers) slides.pptx (warm brown theme) review_game.html (interactive) anki_flashcards.tsv (spaced repetition) kahoot_quiz.csv (import to Kahoot) narration.txt (TTS-ready audio script) study_guide.txt (review sheet) lesson.imscc (Canvas/Moodle import)
Features
- Voice learning -- ingests your files, extracts your teaching style, generates content that sounds like you
- Quality gate with auto-retry -- 12 pedagogical checks (Bloom's progression, primary source depth, differentiation specificity, diversity audit). Bad output gets rejected and regenerated with feedback.
- Your images, your slides -- extracts maps, cartoons, diagrams from your PPTX files. Vision model filters images for quality before embedding.
- Structured jigsaws + creative activities -- timed rotations, graphic organizers, role plays, debates, podcast scripts, gallery walks, mock trials
- Multi-day project arcs -- 5-day projects with choice boards, research databases, rubrics, debate prep sheets, culminating performances
- Export everywhere -- Teacher DOCX, Student DOCX, PPTX, Anki flashcards, Kahoot quizzes, Common Cartridge (.imscc), audio narration, study guides
- Autonomous scheduling -- Ed auto-generates missing lessons at 6am, drafts next week's plans Sunday night, scans standards coverage, and improves its own quality
- Chrome extension -- highlight text on any webpage, right-click, generate a lesson
- Real-time classroom -- live WebSocket sessions with slide control, timers, and polls. Students connect via class code.
- Exit ticket auto-grading -- keyword analysis + formative feedback with sentence starters
- Adaptive feedback loop -- exit ticket results feed into next lesson (reteach/extend recommendations)
- Multi-provider -- Anthropic, OpenAI, Google Gemini, Ollama (local or cloud), OpenRouter
- 51+ agent tools, interactive HTML games, animated videos via Manim
- 50-state standards alignment -- CCSS, NGSS, C3, and state-specific databases
- Telegram bot with file delivery, Google Drive integration, MCP server
- Central approval policy -- every tool classified by risk level, sensitive actions require teacher confirmation
- Data stays local -- no accounts, no telemetry, no data collection
How voice learning works
When you run clawed ingest, Ed reads through your existing
lesson plans, worksheets, assessments, and slides. It extracts:
- Your vocabulary choices and phrasing patterns
- How you structure lessons (warm-ups, activities, exits)
- Your question framing style (Socratic, inquiry-based, direct, etc.)
- Scaffolding moves you use repeatedly
- Assessment preferences (multiple choice, DBQs, open response)
This creates a voice persona stored locally on your machine. Every time Ed generates content, it applies that persona so the output matches how you actually teach. A physics teacher who relies on worked examples and checkpoint questions will get different output than a history teacher who leads with primary sources and Socratic seminars.
You can have multiple personas (e.g., one per course). The ingest command can be pointed at different folders to build them.
# ingest AP History files into one persona clawed ingest ./ap-history/ --persona "AP-Rivera" # ingest regular History files into another clawed ingest ./history-10/ --persona "Gen-Rivera" # generate with a specific persona clawed lesson "The New Deal" --persona "AP-Rivera"
AI providers
Claw-ED is BYOK (bring your own key). You pick a provider during setup. Switch anytime. Here's what's available:
| Provider | Models | Notes |
|---|---|---|
| Anthropic | Claude Opus 4.6 (1M ctx), Sonnet 4.6, Haiku 4.5 | Best voice matching and pedagogy. 128K output. Best quality |
| OpenAI | GPT-5.4, o3, GPT-4.1 | Strong reasoning and coding. Reliable output across subjects. |
| Gemini 3.1 Pro, 2.5 Flash | 94.3% GPQA reasoning. Free tier available for trying things out. | |
| Ollama Cloud | GLM-5.1, Gemma 4, Llama 4 Maverick (400B, 10M ctx) | ~$20/mo unlimited. Best value for daily use. Recommended |
| Ollama Local | Gemma 4, Llama 4, GLM-5, DeepSeek V3.2, Qwen 3 | Free. Fully offline. Quality depends on hardware. DeepSeek V3.2 is 90% of GPT-5.4 at 1/50th cost. |
| OpenRouter | All models from all providers | One API key, access to everything. Pay per token. Best for trying multiple models. |
For most teachers, Ollama Cloud at ~$20/month hits the sweet spot — unlimited access to GLM-5.1, Gemma 4, and Llama 4. For the highest quality voice matching, use Anthropic Claude Opus 4.6 (1M context window, 128K output). Google Gemini has a free tier for trying things out. For fully offline use, install Ollama locally with DeepSeek V3.2 or Gemma 4.
Commands
clawed # chat with Ed clawed lesson "Topic" -g 8 -s "US History" # daily lesson (9+ files) clawed unit "Topic" -g 9 -w 3 # multi-week unit clawed year-map "US History" -g 8 # 180-day curriculum plan clawed assess "Topic" --type crq # CRQ, DBQ, quiz, rubric clawed game create "Topic" -g 8 # interactive HTML game clawed simulate create "Topic" # interactive simulation clawed ingest ~/Documents/Lessons/ # teach it your style clawed differentiate -l lesson.json # IEP/504/ELL mods clawed bot # start Telegram bot clawed serve # start web dashboard + API clawed mcp-server # MCP for Claude Code clawed schedule list # view autonomous tasks clawed drive auth # connect Google Drive clawed setup # re-run setup wizard
Most commands accept --grade, --standards,
--persona, and --format flags. Run
clawed <command> --help for details.
Contributing
Claw-ED is built in public. Contributions are welcome.
Getting started
git clone https://github.com/SirhanMacx/Claw-ED.git cd Claw-ED pip install -e ".[dev]" pytest
Where to help
- Issues -- bug reports and feature requests
- Discussions -- questions, ideas, show and tell
- Good first issues -- if you're looking for something to pick up
Things we'd especially appreciate
- Standards databases for non-US curricula (UK, Canada, Australia, etc.)
- File format support (more ingestion parsers)
- Voice learning improvements
- Quality gate refinements
- Documentation and examples
- Testing with different AI providers and models