Setup multiple ollamas in a local lan

[rustic vessel]

Hi i dont get it. I even tried to set this up with openrouter/auto and i cant get it to work. The remote pc and the mac studio dont respond to openclaw. I can get them to work via curl i terminal but not with openclaw.

1. Orchestrator (Mac mini M4, 16GB)

   • Ollama: Gemma4:E2B
   • Role: Task decomposition, routing, coordination
   • Endpoint: http://localhost:11434 (Ollama default)

2. Worker Alpha (PC with RTX 5080)

   • Llama.cpp: http://192.168.3.3:8080
   • Model: Qwopus-GLM-18B-Healed-Q4_K_M
   • Role: Fast inference, web scraping, initial analysis

3. Worker Beta (Mac Studio, 64GB RAM)

   • Llama.cpp: http://192.168.3.120:8080
   • Model: Qwopus3.5-27B-Q8_0
   • Role: Deep analysis, synthesis, complex reasoning

This would be my setup, any general hints how to set this up? Thx!

[shadow solar]

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

[winged widget]

Short version: openrouter/auto will not make OpenClaw talk to your LAN boxes. Those 192.168.x.x servers need to be configured as their own providers/models.

Big gotchas here:

• localhost means “the machine running the OpenClaw gateway”, not the machine you happen to be typing on.
• Ollama in OpenClaw should use the native URL, for example http://127.0.0.1:11434 and not /v1.
• llama.cpp/OpenAI-compatible servers should usually be configured with a base URL ending in /v1, for example http://192.168.3.3:8080/v1.
• If curl works but OpenClaw does not, test with openclaw infer model run first. Full agent turns are much heavier than tiny curl probes.

What I’d do:

1. Get just the local Ollama model working.
2. Add Worker Alpha as a custom OpenAI-compatible provider.
3. Add Worker Beta the same way.
4. Only then add fallbacks or multi-agent routing.

Also, this is a very real local-model footgun: some backends answer small /v1/chat/completions calls fine, but fail on normal OpenClaw agent turns. If that happens, try:

• compat.requiresStringContent: true
• compat.supportsTools: false
• agents.defaults.experimental.localModelLean: true

If what you want is separate named “orchestrator / worker” identities, that’s Multi-Agent. If you just want multiple backends available to one assistant, keep one agent and add multiple providers.

Docs:

• https://docs.openclaw.ai/providers/ollama
• https://docs.openclaw.ai/start/onboarding-overview
• https://docs.openclaw.ai/gateway/local-models
• https://docs.openclaw.ai/gateway/troubleshooting#local-openai-compatible-backend-passes-direct-probes-but-agent-runs-fail
• https://docs.openclaw.ai/concepts/multi-agent

Please paste these next, especially the version: