What is Ollama MCP Server?
Ollama MCP Server, from the GitHub user rawveg (Tim Green), wraps the official Ollama JavaScript library so an MCP client such as Claude Desktop or Cline can run a local Ollama install. It lists installed and running models, shows model details, pulls, copies, creates, deletes and pushes models, and calls generate, chat and embeddings. With an Ollama Cloud API key it also reaches cloud-hosted models and two web tools, search and fetch.
It has about 178 GitHub stars and roughly 920 npm downloads a week, is written in TypeScript and licensed AGPL-3.0. The last npm release, 2.1.0, came out on 10 November 2025; the only commit since is a dependency security update merged on 13 August 2026.
How to install Ollama MCP Server
npx -y ollama-mcpIs Ollama MCP Server worth it?
This is an administration server, not a delegation layer. Its 13 tools map almost one to one onto the Ollama API (the README says 14, but the published 2.1.0 package ships 13), so an agent can do anything you could do with the ollama command: see what is installed and loaded, pull a new model, build one from a Modelfile, copy it, delete it, or push it to the ollama.com registry. Generate, chat with tool calling and embed cover the inference side.
The security picture follows from Ollama itself. The server speaks stdio only and opens no port of its own, but it talks to Ollama's HTTP API, which by default listens on 127.0.0.1 port 11434 and, per Ollama's docs, needs no authentication locally. Anything that can reach that port can do what this server does. OLLAMA_HOST accepts any address, including another machine or https://ollama.com, with no loopback check. ollama_delete removes a model from disk with no confirmation step in the server, and with OLLAMA_API_KEY set, ollama_push can upload a model under your account. Whether those calls need approval is left entirely to your MCP client.
The Ollama Cloud tools add reach and risk in equal measure. ollama_web_search and ollama_web_fetch send queries to https://ollama.com/api and return page text to the agent, which is a prompt-injection path like any fetch tool; they retry up to 3 times on rate limits and 5xx errors with a 30-second timeout per request.
Maintenance is the weaker point. There has been no npm release in a year, and pull requests that fix Windows tool loading and forward cancellation to Ollama (#23, #24 and #28) sit unmerged. Against mcp-ollama from ~alter, a much newer Ollama server, the trade is clear: this one manages the whole install and reaches the cloud, while mcp-ollama exposes ten narrow task tools, refuses non-local hosts and cannot delete anything.
Where it earns its place
A developer testing local models for a side project
Compare three small models on the same prompt without leaving the chat
Ask the agent to pull each model, run the prompt through ollama_generate and report output and timing, then delete the losers. Free on local hardware; approve each delete in the client.
An engineer building a small RAG prototype on a laptop
Embed a folder of notes with a local embedding model
The agent pulls nomic-embed-text and sends batches through ollama_embed, so nothing leaves the machine and no embedding API is billed. Writing the vectors to a store is a job for another tool.
A solo consultant with an Ollama Cloud account
Let a local model answer questions that need a current web page
Run in hybrid mode with OLLAMA_HOST on 127.0.0.1 and OLLAMA_API_KEY set; the agent searches and fetches through Ollama Cloud and summarises locally. Web usage counts against the Ollama account.
Before you commit
- Setup
- Add npx -y ollama-mcp to the Claude Desktop or Cline config, or npm install -g ollama-mcp. Needs Node.js and a running Ollama; a few minutes to the first tool call. Documented for Claude Desktop and Cline.
- Integrations
- 13 tools: ollama_list, ollama_show, ollama_pull, ollama_push, ollama_copy, ollama_delete, ollama_create, ollama_ps, ollama_generate, ollama_chat, ollama_embed, ollama_web_search and ollama_web_fetch. Five of them (pull, push, copy, create and delete) change state on disk or in the registry.
- Data and privacy
- Stdio only; no port opened by the server. It calls whatever OLLAMA_HOST names (default http://127.0.0.1:11434), and Ollama's local API has no authentication. With OLLAMA_API_KEY set, the key is sent as a bearer token and web search and fetch go to ollama.com. Fetched pages are untrusted text. No published security advisories.
- Cost at scale
- Free locally; cost is your hardware. Cloud models and web tools draw on an Ollama Cloud plan; the server adds nothing on top.
- Limits
- README tool count (14) disagrees with the package (13). Windows tool loading and request cancellation fixes are in unmerged pull requests. No confirmation or allow-list for destructive tools.
What’s good
- Covers nearly the whole Ollama API, including create, copy and push
- Local and Ollama Cloud models from one server
- npx install with no build step
- Retry with Retry-After support on the web tools
What’s not
- Agent can delete or push models with no server-side confirmation
- No loopback check on OLLAMA_HOST
- Last npm release November 2025; fixes wait in open pull requests
- AGPL-3.0 is awkward for some company policies
Best for
Developers who run Ollama on their own laptop and want an agent to manage models and call them, with tool approvals switched on in the client.
Ollama MCP Server alternatives
Ollama MCP Server: common questions
- Is Ollama MCP Server worth it?
- A complete remote control for Ollama that an agent can use to delete or publish your models as easily as to chat with them, so it suits a developer's own machine better than anything shared. We rate it 3.6 out of 5.
- How much does Ollama MCP Server cost?
- Free and open source (AGPL-3.0). Local Ollama is free; cloud models and the web tools need an Ollama account and API key, billed under Ollama's own plans.
- What are the downsides of Ollama MCP Server?
- Agent can delete or push models with no server-side confirmation; no loopback check on OLLAMA_HOST; last npm release November 2025; fixes wait in open pull requests; aGPL-3.0 is awkward for some company policies.
- Who is Ollama MCP Server best for?
- Developers who run Ollama on their own laptop and want an agent to manage models and call them, with tool approvals switched on in the client.
- How much does Ollama MCP Server cost for a team or in production?
- Free locally; cost is your hardware. Cloud models and web tools draw on an Ollama Cloud plan; the server adds nothing on top.
- Is Ollama MCP Server safe for company data?
- Stdio only; no port opened by the server. It calls whatever OLLAMA_HOST names (default http://127.0.0.1:11434), and Ollama's local API has no authentication. With OLLAMA_API_KEY set, the key is sent as a bearer token and web search and fetch go to ollama.com. Fetched pages are untrusted text. No published security advisories.
- What are the best Ollama MCP Server alternatives?
- The closest alternatives we have reviewed are mcp-ollama (~alter).
Answers checked Nov 11, 2026.