How to Use Free LLM Models with OpenHuman — Complete config.toml Setup

Alternative free models (if you hit rate limits)

Why OpenRouter as the inference_url Makes Sense

OpenHuman's power comes from routing different tasks to different models. If you point inference_url at a single-model provider (e.g., Google AI Studio for just Gemini), you give up that routing. OpenRouter is the only free backend that lets you:

• Use one API key for hundreds of models
• Route reasoning → Nemotron 3 Super, coding → DeepSeek V4 Flash, vision → Lyria 3 Pro
• Stay under rate limits by spreading load across different models
• No credit card required for free models

If you prefer not using OpenRouter, Groq is a solid single-provider alternative — Llama 4 Maverick handles most tasks reasonably, and Groq's LPU hardware makes it the fastest free option.

Step 4: Verify It Works

After editing config.toml:

1. Restart OpenHuman
2. Open Settings → AI & Skills → check that your model appears in the status panel
3. Send a test message — if the response comes back, the custom backend is working

If you see connection errors, double-check that your inference_url includes the full path (/v1) and that your API key is valid. Test the key directly:

curl -s https://openrouter.ai/api/v1/models \\
  -H "Authorization: Bearer YOUR_KEY" | head -20

Free + OpenHuman: What You Get vs. Paid

The free setup gives you the core OpenHuman experience — desktop agent, memory tree, 118+ integrations, voice — without the monthly cost. The trade-off: no access to Claude or GPT-5 quality models, and you manage which model does what. For many developers, that's a good trade.

FAQ

Can I use multiple free providers at once?

OpenHuman supports a single inference_url. If you want to mix providers (e.g., Groq for speed + NVIDIA for reasoning), use OpenRouter — it aggregates all of them behind one endpoint.

Does the memory tree work with a free API?

Yes. The memory tree and Obsidian wiki run locally on your machine using SQLite. They call the LLM for summarization — which goes through your custom inference_url. A fast model like Gemini 2.5 Flash handles summarization well.

What if I need Claude or GPT-5 for a specific task?

You can keep the paid subscription for the main orchestrator and set up specific teams in config.toml to use free models for less critical tasks. This hybrid approach gives you the best of both worlds.

Is this officially supported by OpenHuman?

Yes — inference_url is a documented config option in OpenHuman's schema. It's designed for self-hosted/alternative backends. The OpenHuman backend (auth, integrations, OAuth proxy) still runs on their infrastructure; only inference is redirected.