What it does
The server exposes tools that Claude Code can call to run routine work locally: offload_work for code generation and refactoring, compress_context for summarising long text, and batch_offload for running several tasks in sequence or in parallel. Auto_setup recommends a model based on detected RAM, then downloads and preloads it. Other tools cover model recommendation, model pulling and preloading, listing loaded models, runtime selector settings, metrics in JSON or Prometheus format, and cost_dashboard, which tracks cumulative savings and model usage. The publisher reports 736 tests and 97% coverage, and describes the project as a beta for personal use.
It suits boilerplate generation, unit test writing and text summarisation, which the publisher rates as good on local 7B-class models. Simple refactoring is rated as practical, while architecture design and complex debugging are rated as better left to Cloud models. The publisher estimates that if about 40% of API requests are routine, monthly savings could be around $50 to $80. If Ollama is unavailable or slow, the server can switch to the Cloud API. Inputs and outputs pass through prompt injection detection across 5 categories and 20 patterns, and output sanitisation replaces 11 kinds of secrets, such as API keys, passwords and JWTs, with [REDACTED].
Setup requires Node.js 20 or later and Ollama, with Ollama running. The user adds the server to a project-level .mcp.json file, then asks Claude Code to run auto_setup, which downloads a model of roughly 4 GB for typical RAM sizes. Running a model locally needs enough memory for the chosen tier, and the repository also documents building from source. Supported platforms are macOS, Linux and Windows, provided Ollama runs there. The project is licensed under Apache License 2.0.
Limitations and requirements
Requires Node.js 20 or later and an installed Ollama. Model downloads are recommended by RAM tier; the typical download is about 4 GB and larger tiers are much bigger. The publisher describes the project as beta for personal use. The publisher states that local 7B-class output is below Claude's quality, and that architecture design and complex debugging should go to the Cloud. Savings figures are the publisher's estimate and depend on the share of routine tasks.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
io.github.BlackFoil/claude-token-saver-mcp
MCP server that offloads routine coding tasks to local LLM (Ollama) to save Claude API tokens
v0.3.2As shown on registry.modelcontextprotocol.io when we last read it