Claude Token Saver MCP

What claude-token-saver-mcp MCP server does

The claude-token-saver-mcp MCP server connects Claude Code with Ollama so routine work can run on locally hosted language models instead of using a cloud API. Its intended workloads include boilerplate generation, unit test creation, text summarization, and straightforward refactoring. More demanding tasks, such as architecture design or complex debugging, remain better suited to a cloud model.

The server can inspect available system memory, recommend an Ollama model tier, download that model, and load it for use. The documented tiers are Light for systems below 16 GB of RAM, Standard for 16–48 GB, and Ultra above 48 GB. The corresponding examples are phi4:latest, qwen2.5-coder:7b, and qwen2.5-coder:32b.

How it works

Claude Code calls the MCP server over its configured MCP connection. The server sends local work to Ollama over HTTP, where the selected model processes it. If Ollama is unavailable, too slow, or the local result is not appropriate, cloud fallback can be enabled.

The claude-token-saver-mcp MCP server includes input checks for prompt injection patterns and sanitizes sensitive-looking output. Documented redaction patterns include API keys, passwords, and JWTs. The repository describes local processing as having no external data transmission for these local operations.

Execution and cost information are tracked so the server can report estimated savings and model usage. The README gives a monthly savings estimate of roughly $50–80 when routine work represents about 40% of requests; this is an estimate rather than a guaranteed result.

Setup and configuration

Install Node.js 20 or newer and Ollama, then start Ollama with ollama serve. Add the server to a project-level .mcp.json file:

{
  "mcpServers": {
    "token-saver": {
      "command": "npx",
      "args": ["-y", "claude-token-saver-mcp"]
    }
  }
}

After starting Claude Code, ask it to configure a local coding model. The setup flow can recommend a model based on available RAM, download it, and preload it. A source build is also documented: clone the repository, run npm ci && npm run build, and configure Node to execute dist/server.js.

The project documents support for macOS, Linux, and Windows where Ollama runs.

Tools and capabilities

The documented primary tools are:

• offload_work for local code generation and refactoring.
• compress_context for summarizing long text locally.
• auto_setup for model recommendation, download, and preload.
• batch_offload for processing multiple tasks sequentially or in parallel.
• cost_dashboard for cumulative savings and model-use statistics.

Additional tools expose server metrics, model recommendations, model downloads, model preloading, loaded-model listings, and runtime model-selector configuration. These are named get_metrics, recommend_model, pull_model, preload_model, list_loaded_models, and configure_model_selector.

Limitations and notes

The project is marked beta and described as intended for personal use. Local 7B-class models may produce lower-quality results than Claude, especially for architectural decisions and difficult debugging. Local models also require disk space and memory; the documented downloads range from about 2.5 GB to 18 GB. The claude-token-saver-mcp MCP server is therefore most suitable for repeatable, bounded tasks where local output quality is acceptable.