How can I configure OpenHuman to use a local LLM with Ollama?

[nayar-900]

I'm trying to understand the recommended way to configure OpenHuman to use a locally hosted LLM through Ollama instead of a cloud-based model provider.

My goal is to keep the setup as local as possible, including the model and OpenHuman's memory, while still being able to use the normal agent and memory features.

Could someone clarify:

• What is the recommended Ollama configuration for OpenHuman?
• Which model/provider settings need to be changed?
• Does OpenHuman automatically detect an Ollama installation, or does it need an explicit configuration?
• Are there any recommended models for a reasonable balance between response quality and local resource usage?
• Are there any limitations when using Ollama compared with supported cloud providers?

I'm particularly interested in the recommended setup for a local development machine rather than a production deployment.

Thanks!

[meyshad]

Yes. Ollama can handle the chat/agent workloads and the Memory Tree locally, but the routing is deliberately opt-in.

For a development machine, the least error-prone setup is Settings → AI & Skills → Local AI. Start Ollama on its default address (http://127.0.0.1:11434), choose a preset, and let OpenHuman pull the required models. OpenHuman can probe a running Ollama server and list its models, but detection alone does not move any workload off the cloud; you still need to enable the local runtime and select the routes.

If you prefer to edit ~/.openhuman/config.toml, a fully local starting point for a machine with about 16 GB of RAM is:

# Workload routes are top-level config keys.
chat_provider = "ollama:gemma3:4b-it-qat"
reasoning_provider = "ollama:gemma3:4b-it-qat"
agentic_provider = "ollama:gemma3:4b-it-qat"
coding_provider = "ollama:gemma3:4b-it-qat"
memory_provider = "ollama:gemma3:4b-it-qat"
embeddings_provider = "ollama:bge-m3"
heartbeat_provider = "ollama:gemma3:4b-it-qat"
learning_provider = "ollama:gemma3:4b-it-qat"
subconscious_provider = "ollama:gemma3:4b-it-qat"

[local_ai]
runtime_enabled = true
opt_in_confirmed = true
provider = "ollama"
base_url = "http://127.0.0.1:11434"
chat_model_id = "gemma3:4b-it-qat"
vision_model_id = "gemma3:4b-it-qat"
embedding_model_id = "bge-m3"

Pull the two models first:

ollama pull gemma3:4b-it-qat
ollama pull bge-m3

gemma3:4b-it-qat is a useful balance on a 16 GB machine and can handle both text and vision. On an 8 GB machine I would start with gemma3:1b-it-qat for chat and leave vision_model_id empty. Keep bge-m3 for embeddings: the Memory Tree expects 1024-dimensional vectors, so all-minilm and nomic-embed-text are not drop-in replacements.

For Ollama on another host or inside a container, set OPENHUMAN_OLLAMA_BASE_URL to the reachable address instead of localhost.

The main limitations are model quality, speed, RAM use, and vision support. Also, if Ollama becomes unreachable, ordinary routing can fall back to a remote provider. Enable Privacy Mode if local-only inference must be enforced. Sign-in, OAuth integrations, billing, and hosted features may still contact the OpenHuman backend; local Ollama only makes the configured inference workloads local.