Run Claude Code on your own GPU

TurboLLM's Anthropic-compatible endpoint means Claude Code can run against whatever model you've loaded — no cloud key, fully offline.

Get started (one command)

From any terminal, with TurboLLM installed:

$ turbollm launch claude

That's it. TurboLLM points Claude Code's ANTHROPIC_BASE_URL at your local daemon, pins ANTHROPIC_MODEL to whatever model is loaded — so the status line, /status, and context-window sizing all reflect the real model — then execs claude. Any extra arguments are forwarded through.

If no model is loaded, it auto-loads your last-used one (or the first in your library). To pick a specific one:

turbollm launch claude --model qwen2.5-coder-14b

What it can do

Everything Claude Code normally does — reading and editing files, running commands, multi-step agentic tasks — except the model runs on your own GPU. Nothing about your code, prompts, or keys leaves the machine.

Inside Claude Code, turn on Auto Model Swap in Settings → Gateway so its /model picker can list and switch between your local models mid-session.

Recommended models

Coding agents lean hard on tool-calling and long context. Good local picks by GPU size:

8 GB: Qwen2.5-Coder 7B — still the best dense coder that fits a small card.
16 GB: Qwen3-Coder-Next — the newest agentic coding MoE; its experts run from system RAM, so auto-fit fits it on a 16 GB card.
24 GB+: Qwen3-Coder-Next — a higher quant and lighter offload make it noticeably stronger here.

Give it room for context

Agentic sessions fill the context window fast. Load the model with a generous context length — TurboLLM's auto-tune shows a VRAM-fit verdict so you can see how much context fits before you commit.

Manual setup (no magic)

The launcher just sets environment variables — there's nothing it does you can't do by hand. Start TurboLLM, then point Claude Code at it:

# TurboLLM running on http://localhost:6996
export ANTHROPIC_BASE_URL=http://localhost:6996
export ANTHROPIC_AUTH_TOKEN=<your TurboLLM API key, or any non-empty value if auth is off>
export ANTHROPIC_MODEL=<the model you loaded>

claude

On Windows PowerShell, use $env:ANTHROPIC_BASE_URL = "http://localhost:6996" (and so on) instead of export. Get a key from Developer → API Keys → Create — required only if you've turned on Require API key under Settings → Network (e.g. for LAN sharing).

claude not installed?

The launcher tells you how. Or install it yourself: npm install -g @anthropic-ai/claude-code, then re-run turbollm launch claude.

Back← All integrations Nextopencode →