Troubleshooting

LM Studio Issues

API returns 400

The model isn't loaded or is too small to produce valid JSON output.

Fix: Load a recommended model (qwen3-4b, mistral-7b, llama-3.1-8b) in LM Studio's Local Server tab, then click Load Model.

Avoid models under 3B — they can't reliably produce the JSON format GuardClaw needs.

JSON parse failure

Failed to parse response: Unexpected token '<'

The model output <think> tags instead of JSON. Use a model with SAFEGUARD_BACKEND=lmstudio and temperature 0.1.

Or switch to the fallback backend temporarily:

guardclaw config set SAFEGUARD_BACKEND fallback

"No models loaded"

In LM Studio: click the model name → Load Model → wait for green Loaded status.

Analysis is slow (5-10s per call)

• Enable GPU acceleration in LM Studio settings (max GPU layers)
• Use a smaller model: phi-3-mini (3B) is fast with good accuracy
• Or use a cloud backend for speed

---

Server Issues

Server won't start

guardclaw status    # check if already running
guardclaw stop      # stop any stale process
guardclaw start

Dashboard not loading

Make sure the server is running:

guardclaw status

Then open localhost:3002 manually.

Hooks not working (Claude Code)

Reinstall hooks:

node scripts/install-claude-code.js

Then restart Claude Code. Hooks only activate after a restart.

---

Config Issues

Changes not applying

All config changes apply immediately via hot-reload. If a change doesn't seem to take effect, check the server is running:

guardclaw status
guardclaw config show

Wrong backend being used

guardclaw config show   # verify SAFEGUARD_BACKEND value

Config is stored at ~/.guardclaw/.env.

---

Getting Help

Open an issue at github.com/TobyGE/GuardClaw/issues with:

• Output of guardclaw status
• Your SAFEGUARD_BACKEND and model name
• The full error message