Inference and library administration sit in the same tool list, so an agent that needs a model it does not have can pull it, check what is loaded, run the task and delete it afterwards inside one session. The two web tools are the only ones that reach off the machine, and they are the only ones that need a key, which keeps the boundary easy to reason about. AGPL-3.0 is the one detail to settle before this goes inside something you ship.
A server over a local Ollama instance covering inference, embeddings and the model library, plus Ollama's hosted web search and fetch.
- Multi-turn chat with system messages, tool calling and generation options, or a single-turn completion when the conversation shape is not needed
- Embeddings returned as numerical vectors for text you pass in
- The local library administered end to end: list what is installed, show a model's modelfile, parameters and architecture, pull, push, copy and delete
- Which models are loaded in memory right now, separately from which are on disk
- New models created from an existing one with their own system prompt, template and license
- Web search and page fetch through Ollama's hosted APIs, returning a page's title, content and links to ground an answer in current material
An Ollama instance the server can reach, with the models you intend to call already pulled into it. OLLAMA_API_KEY is needed by ollama_web_search and ollama_web_fetch, which call Ollama's hosted APIs; nothing else in the list asks for it. Licensed AGPL-3.0.
One command — npm install -g ollama-mcp