How to point coding agents at a local LLM with Unsloth Start

You already have a coding agent you like. Unsloth Start is how you point it at a local model without rewriting that agent’s provider config. After Unsloth is installed and a model is loaded, you launch from the project directory with unsloth start plus the agent name. Unsloth configures the endpoint, API key, provider, model, and context length for that launch. The whole workflow can run offline on your own hardware.

Unsloth uses temporary or session-scoped provider configuration. It does not add an Unsloth provider to the agent’s normal configuration files. Claude Code, Codex, OpenCode, Hermes, OpenClaw, DeepSeek Harness, and Pi keep their own harnesses. You just start them against Unsloth.

Launch from the project folder

Make sure you have Unsloth installed. Open Unsloth, load a model, go to your project folder, and run:

unsloth start claude

Replace claude with any supported agent:

Claude Code: unsloth start claude
OpenAI Codex: unsloth start codex
DeepSeek Harness: unsloth start dsh
OpenCode: unsloth start opencode
Hermes Agent: unsloth start hermes
OpenClaw: unsloth start openclaw
Pi Agent: unsloth start pi

Without --model, Unsloth uses the first model it reports. If you set no sampling flags, it selects recommended settings for that model, including context length and temperature.

Load a model, or spin a temporary server

You can select and load a model in the same command:

unsloth start claude \
  --model unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL \
  --context-length 32768 \
  --temp 1.0 \
  --top-p 0.95 \
  --top-k 20 \
  --min-p 0.0 \
  --reasoning-effort medium

The :UD-Q4_K_XL suffix selects the GGUF quant.

On the default local address, passing --model lets unsloth start start a temporary server when Unsloth Studio is not already running. That temporary server stops when the agent exits. If Unsloth is already running, the command connects to it and leaves it running. Use --serve or --no-serve to allow or prevent automatic local server startup.

Load options apply when Unsloth needs to load or reconcile the requested model:

--model, -m — select a model
--gguf-variant — GGUF quantization variant
--context-length, --max-seq-length — requested context length when loading
--load-in-4bit / --no-load-in-4bit — 4-bit loading for non-GGUF Hugging Face models
--tensor-parallel / --no-tensor-parallel — tensor-parallel GGUF loading on multi-GPU systems
--reasoning — on, off, or auto (auto follows the model’s chat template, which usually means on)
--reasoning-effort — effort passed to the chat template, for example medium; levels are per model, so use one that model accepts. Unset keeps the template’s own value

Codex needs GGUF via llama-server. Codex currently requires a GGUF model served through the llama-server backend. The current Codex integration does not use the Unsloth transformers backend. If Codex does not connect, that mismatch is the usual cause.

If no Unsloth server is found, start Unsloth and load a model, or pass --model so Unsloth can start a temporary local server.

Connect to a remote Unsloth server

Set the Unsloth URL and API key before launching:

export UNSLOTH_STUDIO_URL=https://studio.example.com
export UNSLOTH_API_KEY=sk-unsloth-...
unsloth start claude

You can also pass the key with --api-key. For a verified local Unsloth server, unsloth start creates or reuses the API key automatically.

Pass-through flags, `--no-launch`, and `--yolo`

Arguments that are not Unsloth options are passed to the selected agent:

unsloth start claude --continue
unsloth start codex --persist resume --last
unsloth start opencode run --continue "Continue the previous task"
unsloth start pi --persist --continue

Use the agent’s own help command for its complete list of native options.

--no-launch prints the generated environment and command instead of starting the agent:

unsloth start claude --no-launch

That output can contain connection credentials. Do not publish it in logs or screenshots.

--yolo maps to the selected agent’s trust or non-prompting mode. It can reduce approval prompts and allow the agent to run commands without asking first. Only use it in an environment where unrestricted agent actions are acceptable.

Sessions and `--persist`

--persist keeps managed storage. It does not resume a conversation by itself. Also pass the agent’s normal resume command.

Claude Code and OpenCode use their normal session stores, so you do not need --persist:

Claude Code: unsloth start claude --continue
OpenCode: unsloth start opencode run --continue "Continue"

Codex, OpenClaw, Hermes, and Pi use a managed home that is temporary by default. Use --persist from the first launch and again when you return.

Codex

unsloth start codex --persist
unsloth start codex --persist resume --last

OpenClaw — keep the same native session ID:

unsloth start openclaw --persist \
  agent --local --session-id my-session --message "Inspect this repository"

unsloth start openclaw --persist \
  agent --local --session-id my-session --message "Continue"

Hermes Agent

unsloth start hermes --persist --oneshot "Inspect this repository"
unsloth start hermes --persist --continue --oneshot "Continue"

Pi Coding Agent

unsloth start pi --persist --print "Inspect this repository"
unsloth start pi --persist --continue --print "Continue"

If a session was not restored, use --persist on the first and later launches, then pass the agent’s native resume option.

From a project directory with Unsloth installed and a model loaded, run unsloth start claude. If you plan to use Codex, OpenClaw, Hermes, or Pi across sessions, add --persist on the first launch. For Codex, load a GGUF model on llama-server first.

Sources

Unsloth Start