Guide
Use an uncensored model in Crush
Crush connects to unrestricted.sh through its OpenAI-compatible provider, and one command sets it up. You get Qwen3.8 27B Uncensored, a model for coding and chat with 262K context, at $1.00 per million input tokens.
Get an API key
Sign up, then add credit: API keys work after your first top-up of $5 or more. A first top-up of $10 or more adds $5 of bonus credit.
Connect Crush
sh -c "$(curl -fsSL https://unrestricted.sh/setup.sh)"Run it in a terminal and approve the connection in your browser. It creates an API key for this machine, asks which tools to set up, and writes the Crush settings to ~/.config/crush/crushrc, leaving the rest of that file as it was. It needs curl and Node 18 or newer.
Already have a key? Set UNRESTRICTED_API_KEY to it first and the browser step is skipped. Put UNRESTRICTED_TOOLS=crush in front of the command to set up only Crush.
Or set it up by hand
This adds the provider and the model to ~/.config/crush/crushrc. Run it with UNRESTRICTED_API_KEY set and your shell fills in the key. Delete any unrestricted lines an earlier setup left in the file first.
mkdir -p ~/.config/crush
cat >> ~/.config/crush/crushrc <<EOF
provider add unrestricted --type openai-compat --base-url "https://api.unrestricted.sh/v1" --api-key "$UNRESTRICTED_API_KEY"
model add unrestricted/qwen3.8-27b-uncensored --name "Qwen3.8 27B Uncensored" --context-window 262144 --default-max-tokens 32768 --can-reason true
model use unrestricted/qwen3.8-27b-uncensored
EOF
chmod 600 ~/.config/crush/crushrcPick the model
The model use line makes Crush start on Qwen3.8 27B Uncensored. Run crush.
If something goes wrong
- 401 invalid_api_key
- Missing or revoked API key.
- 402 insufficient_funds
- Out of funds.
- 403 api_locked
- Add funds once to unlock API keys.
- 429 concurrent_request_limit
- Too many requests running on this account.
Every error the API returns is listed in the API reference.
What a session costs
Cached input: when your tool resends the start of a conversation, those tokens cost $0.10 per million instead of $1.00. Cached-input pricing applies to the API and coding tools; messages in the web chat pay the full input price.
A coding-agent turn with 60,000 tokens in, 55,000 of them cached, 1,500 out costs $0.015. The pricing page has every price and more examples.