Best balance of capability and privacy. Qwen runs on your Mac's Metal GPU via llama.cpp, supports tool use, and gives Claude Code 32K of context to work with, without sending a byte to any cloud.
Launch ToolPiper. Open Models → Browse. Search for Qwen. Pick the largest GGUF that fits your RAM: ToolPiper shows the predicted footprint per quantization. Qwen2.5-Coder-14B-Instruct-Q4_K_M is a strong default for 16 GB Macs; Qwen2.5-Coder-32B-Instruct-Q4_K_M for 32 GB+.
In ToolPiper → Endpoints → Add Endpoint → llama.cpp (local). Pick the Qwen model you just downloaded. Set the context window to 32 K. Save. ToolPiper warms the model in the background; first request is fast even on a cold start.
Ask ToolPiper in chat to "set up Claude Code", or call the claude_code_install MCP tool. ToolPiper symlinks the bundled claude-tp helper onto your PATH, mints a scoped inference token, and writes the helper's own mode-0600 config file; it doesn't touch ~/.claude/settings.json. Your Qwen endpoint shows up in Claude Code's /model picker automatically. The install is idempotent, so running it again just refreshes the symlink and token.
Open a new terminal window so it picks up the updated PATH. The claude-tp helper handles the rest.
claude-tp
# /model → "Local Qwen"
# Now every prompt runs on your Mac.In ToolPiper, edit the endpoint and bump the context window override. ToolPiper re-warms with the new size. The claude-tp helper rebuilds its per-launch config from ToolPiper's live endpoint list on every run, so the next claude-tp launch just picks it up.
GGUF on Metal GPU, llama.cpp, your filesystem. Nothing leaves the Mac. ToolPiper's resource scheduler evicts cleanly when memory pressure hits.
Qwen2.5-Coder is tool-trained: Bash, Edit, Read, Write all work. Claude Code's full feature set, locally.
Hit a hard problem? Say "switch to my OpenAI endpoint" mid-session. ToolPiper reroutes; the conversation continues with cloud horsepower for that one task.
On 16 GB: Qwen2.5-Coder-14B Q4_K_M. On 32 GB: Qwen2.5-Coder-32B Q4_K_M. ToolPiper's model browser shows predicted memory + tokens/sec on your specific Mac, so you don't have to guess.
For most edits, yes. Claude Code's <code class="font-mono text-xs text-fg-muted">endpoint_recommend</code> tool will suggest a longer-context backend automatically when a conversation overflows; it sees your whole endpoint list and picks the right one.
ToolPiper uses the same llama.cpp engine Ollama wraps. The difference is the platform around it: over 420 MCP tools, browser automation, an Anthropic proxy with provider switching, and resource intelligence. See the <a href="/compare/claude-code-backends" class="text-status-success hover:text-status-success-hover">comparison page</a>.
Yes. ToolPiper's endpoint serves <code class="font-mono text-xs text-fg-muted">/v1/chat/completions</code> (OpenAI-shape) and <code class="font-mono text-xs text-fg-muted">/v1/messages</code> (Anthropic-shape) simultaneously. Same model, both shapes, no extra config.
ToolPiper is a free download. Configure once and Claude Code routes through your Mac.