Ollama (self-hosted)

Connecting Aimogen Pro to a local or remote Ollama server for private, no-per-token-cost inference, including networking, model lists and embeddings.

Ollama runs open models on hardware you control. There are no per-token charges and no prompt data leaves your infrastructure, which makes it the option to reach for when privacy is the constraint.

Prerequisites#

1. Install Ollama on a machine your WordPress server can reach over HTTP.
2. Pull at least one model:

ollama pull llama3.3
ollama pull nomic-embed-text     # if you want local embeddings

1. Make the API reachable from WordPress. By default Ollama binds to 127.0.0.1:11434, which is only reachable from the same machine.

Configure#

Aimogen Pro › Settings › API Keys › Ollama URL

Enter the base URL, for example:

http://127.0.0.1:11434
http://10.0.0.5:11434
https://ollama.example.com

There is no API key. The plugin considers Ollama configured when this URL is non-empty.

After saving, the plugin queries the server and caches the installed models in the aiomatic_ollama_models option, and embedding models in aiomatic_ollama_embedding_models. Only models actually pulled on the server appear in the selectors.

Networking#

Same server as WordPress. http://127.0.0.1:11434 works if PHP can open local sockets. Some managed hosts block this.

Different server. Ollama must listen on a reachable interface:

OLLAMA_HOST=0.0.0.0:11434 ollama serve

A minimal nginx front end with basic auth and TLS is the usual answer.

Models#

The list is whatever you have pulled. The plugin also maintains two static capability lists that matter:

• Vision models (AIMOGEN_MODELS_OLLAMA_VISION) — llama3.2-vision, llava, llava-llama3, bakllava, moondream and similar. Only these accept image input.
• Function-calling models — a large list including llama3.3, llama3.2, llama3.1, qwen2.5, mistral, mixtral, command-r, phi4, deepseek-r1 and many community fine-tunes. Only these work with Chatbot Extensions and agents.

If a model you pulled is not in the plugin capability list, it still works for plain text generation; it simply will not be offered where tools or vision are required.

Embeddings#

Ollama can serve as the embeddings provider, which keeps your indexed content entirely on your own hardware. The plugin recognises a long list of embedding models; the common choices are:

ollama pull nomic-embed-text
ollama pull mxbai-embed-large
ollama pull bge-m3

Configure under Settings › Embeddings.

Token limits#

Ollama responses are capped at AIMOGEN_MAX_OLLAMA_TOKEN_COUNT, 4000 tokens in 2.8.8. For longer articles, generate in sections.

The Ollama Responses API#

Settings › Advanced AI Settings › Enable Usage Of Ollama's Responses API

Off by default. Version 2.8.2 fixed a streaming issue that occurred when Ollama streaming was combined with tool calls; if you see odd streaming behaviour with tools, make sure you are on 2.8.8 or newer and leave this toggle off.

Performance expectations#

Ollama speed is a function of your hardware.

Long generations on modest hardware will exceed PHP max_execution_time. Raise it, use smaller models, or generate in sections. See Recommended server configuration.

Common problems#

No Ollama models in the dropdown WordPress cannot reach the URL. Test from the WordPress server itself:

curl http://127.0.0.1:11434/api/tags

Connection refused from another machine Ollama is bound to localhost. Set OLLAMA_HOST=0.0.0.0:11434.

Timeouts Model too large for the hardware, or PHP limits too low. Try a smaller model first.

Tools never fire The model is not in the function-calling capability list. Use llama3.3, qwen2.5 or another listed model.

Related#

• Hugging Face — the same models, hosted
• Custom OpenAI-compatible APIs — for other self-hosted runtimes
• Security and privacy