Forge - GPU Kernel Optimization

Created By
RightNow-AI4 months ago
Turn slow PyTorch into fast CUDA/Triton kernels. 32 parallel swarm agents optimize your code on real datacenter GPUs (B200, H200, H100, A100) with up to 14x speedup over torch.compile.
Overview

Forge MCP Server

Swarm agents that turn slow PyTorch into fast CUDA/Triton kernels, from any AI coding agent.

What it does

  • Optimize existing kernels - Submit PyTorch code, get back an optimized Triton/CUDA kernel
  • Generate new kernels - Describe an operation, get a production-ready optimized kernel
  • 32 parallel swarm agents - Coder+Judge pairs compete to find the fastest kernel
  • Real GPU benchmarking - Every kernel is tested on datacenter hardware (B200, H200, H100, A100, L40S, T4)
  • Up to 14x faster than torch.compile(mode='max-autotune')
  • One-click auth - Browser-based OAuth, no API keys needed

Quick Start

claude mcp add forge-mcp -- npx -y @rightnow/forge-mcp-server

Tools
┌────────────────┬───────────────────────────────────────────────┐
│      Tool      │                  Description                  │
├────────────────┼───────────────────────────────────────────────┤
│ forge_auth     │ Authenticate with Forge via browser           │
├────────────────┼───────────────────────────────────────────────┤
│ forge_optimize │ Optimize PyTorch code into fast GPU kernels   │
├────────────────┼───────────────────────────────────────────────┤
│ forge_generate │ Generate optimized kernels from a description │
├────────────────┼───────────────────────────────────────────────┤
│ forge_credits  │ Check credit balance                          │
├────────────────┼───────────────────────────────────────────────┤
│ forge_status   │ Check job status                              │
├────────────────┼───────────────────────────────────────────────┤
│ forge_cancel   │ Cancel a running job                          │
├────────────────┼───────────────────────────────────────────────┤
│ forge_sessions │ List past optimization sessions               │
└────────────────┴───────────────────────────────────────────────┘


Pricing

Pay-as-you-go. $15/credit, 25% off at 10+. Free trial included - optimize 1 kernel, no credit card required.

https://rightnowai.co - https://github.com/RightNow-AI/forge-mcp-server - https://www.npmjs.com/package/@rightnow/forge-mcp-server

Server Config

{
  "mcpServers": {
    "forge": {
      "command": "npx",
      "args": [
        "-y",
        "@rightnow/forge-mcp-server"
      ]
    }
  }
}
Project Info
Created At
4 months ago
Updated At
4 months ago
Author Name
RightNow-AI
Star
-
Language
-
License
-
Category

Recommend Servers

View All
Ghl Command
@Elite DCs LLC

GoHighLevel MCP server for Claude. 212 tools across 43 modules, including the only programmatic GHL workflow builder (private API, reverse-engineered), funnel + page editor, form builder, pipeline builder, pre-deploy validator, multi-sub-account switching, bulk operations, and full account export. $97 one-time, lifetime updates. GHL Command gives Claude full programmatic control of GoHighLevel through 212 tools across 43 modules. Built for GoHighLevel agency operators who manage many client sub-accounts and want to onboard new clients in minutes instead of days. Exclusive capabilities (none of the free GHL MCPs have these): - Programmatic workflow builder. Create, edit, clone, publish, and validate complete GHL workflows from a single prompt. GHL's public API has no workflow write endpoints; this uses their internal API (the same one their UI calls). - Funnel + page editor and form builder (also private API). - Pipeline builder, goal event builder, full 57-native-trigger registry. - Pre-deploy validator that catches GHL's silent invalid-ID failure (a common workflow-breaking bug GHL never warns you about). - Multi-sub-account token registry. Switch between any client account mid-conversation; API keys swap automatically. - Bulk operations: tag, update, enroll, delete hundreds of contacts in one command. - Full account export and side-by-side location diff for audit or migration. Works with Claude Desktop App, Claude Code (terminal), and headless on a Linux server or droplet. $97 one-time, 3 machines, no subscription, lifetime updates. 30-day time-back guarantee: save 5+ hours on one real client build or full refund.

a day ago
Tavily Mcp
@tavily-ai

JavaScript
a year ago
Fixmypdf

15 hours ago