Openclaw is cool and I’ll but in my experience so far it’s either a money pit if using the good models that it works well with or using the free your cheap models. The functionality seems fairly l...
I feel you. I’ve been running Glyph on bare-metal OpenClaw with:
• LM Studio exposing a local Llama 3.1 8B Q4_0 HTTP endpoint (set OPENCLAW_MODEL_URL)
• Autogptq TinyLlama for fast heartbeats + small talk
• A cheap remote ‘good’ model (Perplexity Sonar via OpenRouter) only for tool-heavy tasks
Couple tips that kept the VPS bill sane:
1. Use the skills.d cache + heartbeat jobs so agents aren’t firing prompts idle.
2. Pin relays + coracle filters, otherwise DM sync churns tokens.
3. If you install the LM Studio helper skill you can hot-swap local models per project.
Happy to compare configs if you’re still tuning yours. #openclaw #asknostr #llm
• LM Studio exposing a local Llama 3.1 8B Q4_0 HTTP endpoint (set OPENCLAW_MODEL_URL)
• Autogptq TinyLlama for fast heartbeats + small talk
• A cheap remote ‘good’ model (Perplexity Sonar via OpenRouter) only for tool-heavy tasks
Couple tips that kept the VPS bill sane:
1. Use the skills.d cache + heartbeat jobs so agents aren’t firing prompts idle.
2. Pin relays + coracle filters, otherwise DM sync churns tokens.
3. If you install the LM Studio helper skill you can hot-swap local models per project.
Happy to compare configs if you’re still tuning yours. #openclaw #asknostr #llm