# TurboLLM

> TurboLLM is a local-LLM platform published to npm as `turbollm` (run it with `npx turbollm`).
> It runs any llama.cpp-compatible inference engine or community fork, auto-tuned to your GPU, with
> a browser UI and an OpenAI- and Anthropic-compatible API on one port. It reports REAL measured
> tokens/sec and a VRAM-fit verdict per model. It is offline-first: no account, no backend, no
> telemetry, ~2 MB, no Electron, no Python. Requires Node.js 22+.

## Start

- [Install & first run](https://turbollm.dev/docs/getting-started): `npx turbollm` starts a local daemon on http://127.0.0.1:6996 and opens the browser UI.
- [Load your first model](https://turbollm.dev/docs/first-model): reuse existing GGUF folders or browse & download from Hugging Face in-app.
- [Documentation home](https://turbollm.dev/docs)

## Core capabilities

- [Run any engine](https://turbollm.dev/docs/engines): swap in llama.cpp, ik_llama.cpp, TurboQuant, KoboldCpp, llamafile, MLX, vLLM, SGLang, or any community fork — with a hardware-fit verdict per engine.
- [Auto-tuning & speed](https://turbollm.dev/docs/auto-tune): benchmarks your GPU on load, derives fast launch flags (context, -ngl, --n-cpu-moe, KV-cache quant, flash attention, speculative decoding), and shows measured tokens/sec.
- [Chat, skills & agents](https://turbollm.dev/docs/chat): streaming chat, thinking control, live artifacts, agentic tools with an approval gate, editable agent personas, a shared SKILL.md library, and MCP servers.

## API

- [API overview](https://turbollm.dev/docs/api): base URL http://localhost:6996; the gateway loads the requested model on demand.
- [OpenAI-compatible](https://turbollm.dev/docs/api/openai): GET /v1/models, POST /v1/chat/completions, POST /v1/embeddings, GET /v1/health at http://localhost:6996/v1.
- [Anthropic-compatible](https://turbollm.dev/docs/api/anthropic): POST /v1/messages with tool use and streaming — this is what runs Claude Code locally.

## Integrations (point your tools at your own GPU)

- [All integrations](https://turbollm.dev/integrations)
- [Claude Code](https://turbollm.dev/integrations/claude-code): `turbollm launch claude` — one command, no cloud key.
- [Continue (VS Code & JetBrains)](https://turbollm.dev/integrations/continue)
- [Cline](https://turbollm.dev/integrations/cline), [opencode](https://turbollm.dev/integrations/opencode), [Kilo Code](https://turbollm.dev/integrations/kilo)

## Choosing a model

- [Models hub](https://turbollm.dev/models): hand-picked local models by VRAM tier (8 GB / 16 GB / 24 GB+), with capability badges and Hugging Face repos.
- [Quantization explained](https://turbollm.dev/models/quantization): what Q4_K_M, Q8_0, and f16 mean, and which quant fits your GPU.

## Reference

- [CLI reference](https://turbollm.dev/docs/cli)
- [Configuration & data](https://turbollm.dev/docs/configuration): everything lives under ~/.turbollm/.
- [Privacy & offline](https://turbollm.dev/docs/privacy)
- [GitHub](https://github.com/mohitsoni48/TurboLLM) · [npm](https://www.npmjs.com/package/turbollm) · [Discord](https://discord.gg/v6kRbV7nC)