OpenRouter
One API key can route to many hosted models. It is the simplest path for teachers who want strong cloud models without wiring every provider separately.
Create an OpenRouter key
Free Mac app · Open source · Bring your own model
Claw-ED is an AI-agent harness for teachers.
It runs a teacher-focused agent on your Mac, gives it a controlled set of tools, and ships with prebuilt skills for the work teachers actually need done: planning, grading support, local file organization, differentiation, review materials, and curriculum search.
Mac download coming soon
View the open-source project
The Mac build is in prototype testing. Public download will open after the Mac app and companion iOS path have been tested together.
Harness
The Mac app wraps a local agent, not a single prompt box.
Provider choice
Use OpenRouter, Ollama, Anthropic, OpenAI, Google, or local Ollama.
Teacher skills
Education tasks are packaged as reusable agent skills.
Permission prompts
Risky actions ask first, and shell commands are treated carefully.
What a harness means
A general-purpose local agent, tuned for school work.
Claw-ED is not just a lesson generator. It is the local runtime around an agent: workspace access, model configuration, skill loading, tool-risk labels, approvals, file outputs, and teacher-facing UI.
The agent can read approved workspace context, call a model provider, decide which skill or tool to use, ask before higher-risk actions, and write artifacts back to the Mac. Lesson bundles are one skill. The bigger idea is a teacher assistant that can loop through real work.
Choose your model
Bring your own API key from the provider you trust.
Claw-ED does not sell model credits. You connect the model account you already approve.
Ollama local
Install Ollama on the Mac and run a local model. This keeps model calls on the machine, but it requires more technical comfort and enough Mac hardware.
Download Ollama for MacOllama Cloud or other providers
Anthropic, OpenAI, Google, Ollama Cloud, and other cloud providers can be used when a district already has an approved account or data agreement.
Read Ollama API docsSetup
Get a key, save it on the Mac, pick a model.
The in-app setup screen walks through this, and the CLI commands are direct for IT teams.
OpenRouter
clawed config set-key openrouter YOUR_OPENROUTER_KEY
clawed config set-model openrouter --model minimax/minimax-m3Ollama running locally
ollama pull llama3.2
clawed config set-model ollama --model llama3.2
Keys are saved on the Mac. Claw-ED tries the operating-system keychain
first, then falls back to a local secrets file under ~/.eduagent/
if the keychain is unavailable.
Skills and permissions
The agent can act, but it does not get a blank check.
Action
How Claw-ED treats it
Read approved workspace files
Low risk; visible to the local agent context.
Write a lesson, handout, CSV, or report
Local write; approval depends on path and overwrite risk.
Run a shell command
Always treated as sensitive and shown before execution.
Use a cloud model provider
Prompt/context is sent to that provider under its policies.
Trust boundary
The agent is local. Cloud models are still cloud services.
Claw-ED itself runs locally on the Mac. The Mac app talks to a local sidecar over loopback, and files are read or written on the Mac unless the teacher approves another action.
If you configure OpenRouter, Ollama Cloud, Anthropic, OpenAI, Google, or another hosted model, Claw-ED sends the prompt and relevant context to that cloud service. With OpenRouter, requests may also be processed by the downstream model provider selected by OpenRouter or your routing settings. Those provider terms, retention policies, and district agreements matter.
If your priority is keeping model requests off cloud services, run a local Ollama model on the Mac instead. That is the most private path, but it is more technical and model quality depends on the hardware and model you choose.
For schools and districts
A free reason to make the Mac mini the teacher's AI workstation.
Claw-ED is free and open source because the value is in giving educators a capable, inspectable local harness instead of another locked SaaS box.
- Deploy a notarized Mac app with no App Store dependency.
- Use district-approved model providers or local Ollama.
- Keep teacher files on the teacher's Mac unless a provider call is configured.
- Audit skills, tool permissions, and outputs before a pilot.
- Start with planning, differentiation, local materials, and review workflows.
Next skill direction
/goal and /loop are the right next layer.
/goal
Give the agent a concrete educator outcome, a workspace, and success criteria: "build tomorrow's Global 9 lesson packet" or "turn this unit folder into review week materials."
/loop
Let the agent work in controlled cycles: plan, act, check the result, ask for approval when needed, and stop when the goal or budget is met.
Teachers should own their assistant.
Claw-ED is the local harness: open source code, your model provider, your files, your approval decisions. The public Mac download is coming soon after prototype testing.