Train in Silence MCP
Train in Silence estimates the VRAM a fine-tuning job needs and ranks cloud GPU offers by cost and run time.
- 104 ★ on the repository
- Last commit
- MIT
Minimal access
Why this level
- The server reads prices and rents nothing; the job description reaches price providers only when keys are set.
Install
- Works in
Manual install
pip install train-in-silence
claude mcp add tis --scope user -- tis-mcpThe first README option: install the package and register the server in Claude Code. For Claude Desktop add `tis-mcp` to `claude_desktop_config.json`.
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The server reads prices and rents nothing; the job description reaches price providers only when keys are set.
This is third-party code. Review the repository files before installing.
What it does
Train in Silence is an MCP server and CLI that estimates a language-model fine-tuning job (VRAM and compute) and matches it against GPU prices from about a dozen cloud providers. It returns three options: cheapest, fastest and balanced. Every recommendation names its data source: `live:official`, `live:gpuhunt`, `live:gpufinder` or `sample`, and estimated fields are flagged. Vast.ai and RunPod keys are recommended but optional; without them the server falls back to general price aggregators or bundled samples.
Who it is for. ML engineers and developers who fine-tune models and want to pick a GPU by budget and deadline quickly.
Good fit when
- A model must be fine-tuned within 20 hours and the GPU rental options are too many.
- The budget is tight and the cheapest suitable option is needed.
- Before launch, one wants to check whether the VRAM fits the chosen model.
- Vast.ai, RunPod and Lambda prices are compared for one job.
Not a fit when
- Precise calibration from real runs is needed, but the estimation model is fixed.
- Guaranteed fresh prices are required, yet the source may be `sample` data.
- Training runs on your own hardware with no cloud rental.
Example request
I want to run the fine-tune code in this folder and finish within 20 hours. Find the best GPU options across Vast.ai, RunPod and Lambda.MCP
- Transport
- stdio
- Authentication
- not required
Limitations
The project is experimental. The estimation model is fixed and not calibrated against real run times, and upstream provider API changes will need mapping updates. Without keys, prices come from aggregators or samples, so check the source field. The server rents nothing itself. Last commit: 4 August 2026. Prices move in real time, so confirm the final rental on the provider site: the server is an estimate and not a cost guarantee.
How to disable. Run `claude mcp remove tis` and uninstall the package with `pip uninstall train-in-silence`. Remove any provider keys you set from the environment.
FAQ
Are provider keys required?
No. Vast.ai and RunPod keys are recommended for accurate prices; without them the server falls back to the GPUHunt and GPUFinder aggregators or bundled samples.
What does the source field mean?
It shows price freshness: `live:official` comes from the provider, `live:gpuhunt` and `live:gpufinder` from aggregators, and `sample` is bundled data.
How do I use it without an agent?
Through the CLI: `tis recommend request.yaml`. A REST API is also available, started with `uvicorn tis.api.server:app`.
How many options does it return?
A Pareto front: the cheapest, fastest and balanced options.
What is needed besides Python?
Python 3.10 or newer. The server installs as the `train-in-silence` package, and the rest comes from price aggregators.
Discussion
No questions or reviews yet.
Related
MCP serversA skill for Claude Code and Gemini CLI that runs LLM fine-tuning from dataset to GGUF export, on NVIDIA GPUs and Apple Silicon
An MCP server that hands bounded tasks to a local model or a cheap cloud API, while Claude stays the architect and saves tokens
A local model server for Apple Silicon Macs: Qwen 3.8 and Gemma 4 with MTP decoding and OpenAI and Anthropic APIs for Claude Code, Codex and Cursor
agentperf-local measures how fast your machine serves an AI agent on recorded conversations: throughput and latency
Remote MCP server that tells which search, extraction, reranker or retrieval API fits your task best, with confidence intervals.
A local GGUF model server accepts OpenAI-style requests so applications can be tested on hardware you control.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Train in Silence (this card) | no | not required | 104 | |
| Vast.ai CLI | no | 231 | manages rentals on Vast.ai; Train in Silence compares a dozen providers for your job | |
| Runpod MCP Server | yes | OAuth | 87 | manages RunPod resources; Train in Silence estimates VRAM and picks GPUs |
README in short
The document covers the problem, a Claude Code and CLI quickstart, the provider list and the architecture: YAML request, estimator, market aggregator, optimizer and Pareto ranking.