tis-mcp

You want to fine-tune an LLM. You open Vast.ai, RunPod, AWS, etc. -- a dozen tabs, a dozen pricing models, a dozen different ways to describe a GPU. Which option can run your code, and do so more cheaply and quickly? An hour later you're still in a spreadsheet and haven't written a single line of training code.

Train in Silence is the first Task-Aware MCP server for LLM fine-tuning. It doesn't just list prices; it understands your workload. Describe your training job once, and it calculates the required VRAM/FLOPs to return the cheapest, fastest, and most balanced hardware options across a dozen cloud providers -- in seconds.

Quickstart

Option A: Ask Claude Code (recommended)

Install the library and register it as a tool in Claude Code:

pip install train-in-silence
claude mcp add tis --scope user -- tis-mcp

Then just ask in natural language:

> I want to run the fine-tune code in my current directory, and finish it within 20 hours.
  Find me the best GPU options across Vast.ai, RunPod, and Lambda.

Claude Code calls TIS behind the scenes and returns a structured recommendation -- no YAML, no config files, no manual comparison.

Option B: CLI

pip install train-in-silence
tis recommend examples/request.yaml

$ tis recommend examples/request.yaml

  Found 5 viable configurations
  Lowest cost: $4.32 | Fastest runtime: 2.1 hours

  #1 [cheapest]  RunPod 1x A6000 (48 GB)    $4.32 / 6.8 h
  #2 [fastest]   Vast.ai 2x A100 (80 GB)    $9.10 / 2.1 h
  #3 [balanced]  RunPod 1x A100 (80 GB)     $6.40 / 3.2 h
  ...

Note: Output above is illustrative. Actual results depend on live market data.

Related MCP server: Train in Silence

Use It Your Way

Market Providers

TIS aggregates live pricing across a dozen GPU clouds. API keys are optional: if not provided, TIS automatically falls back sequentially to universal live aggregators (GPUHunt/GPUFinder) or bundled sample data.

Every recommendation clearly identifies its Source of Truth (e.g., live:official, live:gpuhunt, live:gpufinder, or sample) so you always know how fresh the data is. -> Provider details

Architecture at a Glance

YAML request -> Estimator -> Market Aggregator -> Optimizer -> Pareto Frontier -> Ranked Output
                  |                |                 |
              VRAM/FLOPs     10+ GPU Clouds    Cost vs. Time

Each recommendation shows where the data came from (live or sample) and flags any estimated fields -- no silent guesswork. -> Architecture deep-dive

Known Limitations

• Estimation model is fixed with no built-in calibration; future versions will calibrate using real runtimes.

• Upstream Provider API schema changes will require synchronized mapping updates.

🚧 Project Status & Contribution

This project is currently in the experimental development stage (Experimental).

• Issues & Suggestions: If you encounter any bugs, inaccurate estimations, or have suggestions for improvement, please feel free to submit a GitHub Issue.

• Contribute: If you'd like to improve the code or supplement hardware metadata, Pull Requests are highly welcome! We look forward to refining this LLM hardware planner with the community.