Unsloth Start launches multiple coding agents locally
danielhanchen · x · 2026-07-20
Unsloth Start can launch multiple agent harnesses and support local workflows with llama.cpp prebuilts and built-in tool-call healing, which the post says improves tool-call accuracy by 50%.
It runs in isolated environments and supports --resume, tmux, and agents such as Claude, Codex, Hermes Agent, OpenClaw, OpenCode, and Pi. The screenshot also shows Unsloth commands like unsloth start claude/codex/hermes/... and examples of Claude Code and Codex running locally on Unsloth models.
In the follow-up reply, the author says Unsloth worked with AMD to add RL, inference, notebooks, and training improvements across 500+ models, including:
- 2x faster runs
- 70% less VRAM for training
- support for RDNA 3/4, Strix Halo, MI300, and 325+ devices/models
- Windows, Linux, and WSL support
- RL vLLM weight sharing and faster execution
- free remote HTTPS access
More from coding & agent
- GitHub review bot hits its PR limit and forces a 39-minute cooldown — DanielLockyer · 2026-07-22
- Max reasoning effort appears to be mobile-only in Codex Remote, not desktop — GabGarrett · 2026-07-22
- A Reddit demo argues online stores should expose carts and pricing through MCP — gelembjuk · 2026-07-22
- Open-source AI SDK provider routes Vercel apps through a local Codex subscription — lgrammel · 2026-07-22
- Codex vs Claude Code: Which Is More Popular? — jxnlco · 2026-07-22
- CodeRabbit uses layers, diagrams and a chat agent to rethink code review — _jaydeepkarale · 2026-07-22