Claude Code, on-device, on your Mac.
Looking to run Claude Code on Apple Intelligence? ToolPiper's Engine switch offers a model only if it can carry Claude Code's tool calls and a 98K window, so the on-device route is a local language model. Free, private, no cloud round-trip, no per-token billing.
Before you start
- Apple Silicon Mac (M1 or later) with enough memory for the model. ToolPiper recommends a model for this Mac on Home.
- Claude Code installed (run
claude --versionto verify). - ToolPiper installed (free DMG from modelpiper.com/download).
Setup
- 1
Choose a local language model
Launch ToolPiper and open Choose a language model on Home. Gemma 4 E4B, Gemma 4 12B and Qwen 3.8 27B each have a trained window of at least 98,304 tokens, which is what Claude Code needs, and Choose another shows how each fits this Mac.
- 2
Turn on Engine in the Claude Code row
Under Connect your coding agent on Home, turn on Engine in the Claude Code row and pick your model. ToolPiper writes a few entries to Claude Code's user settings,
~/.claude/settings.json, and keeps its token in its own folder. Sessions you already have open stay as they are. - 3
Start a new session
Run your normal command. A new session starts on the model you picked, and its first message also loads the model.
shellclaude # In Claude Code, type "/model" to see your ToolPiper models. - 4
Move between models with /model
Claude Code's
/modelpicker now lists every model the switch can offer. Each one compacts at its own number, so sessions on different models run side by side. - 5
Turn it off when you are done
Turn Engine off. New sessions go back to your Anthropic account and the settings file returns as ToolPiper found it. If you picked a model with
/modelin a session, run/modelonce more there.
Why this recipe
Fully private
The model runs on your Mac's Metal GPU through llama.cpp. Your prompts and code never leave the machine.
Your chats stay as they are
Claude Code gets its own entry of the model at 98,304 tokens, and a local chat stays at its own window. The same model in both at once loads twice.
No bill, no quota
There's no token meter. A local model is free on every ToolPiper tier, so run Claude Code as much as you want.
Frequently asked
Why isn't Apple Intelligence on the list?
The switch offers a model only if it can carry tool calls and has a window of at least 98,304 tokens, because Claude Code holds about 33K back for compaction and works through tools on nearly every turn. Apple Intelligence doesn't meet that, so the switch leaves it out and offers the models that do.
Will it work with Claude Code's tool use?
Yes. ToolPiper's Anthropic proxy translates the local model's output into the Anthropic content-block shape Claude Code expects, including tool_use blocks. Bash, Edit, Read, Write all flow through.
Do I need an Anthropic API key?
No. Claude Code uses a ToolPiper-minted token kept in ToolPiper's own folder, not an Anthropic key. The token validates locally against ToolPiper's token store; it never reaches Anthropic.
How do I undo this?
Turn Engine off. ToolPiper revokes the token and puts ~/.claude/settings.json back as it found it, byte for byte when nothing else changed in the file.
Ready to try it?
ToolPiper is a free download. Configure once and Claude Code routes through your Mac.