Forge

Connect MCP clients to Forge and optimize PyTorch code into fast CUDA/Triton kernels using 32 parallel swarm agents with real GPU benchmarking.

typescript

1

GitHub Stars

typescript

Language

4 months ago

First Indexed

2 months ago

Catalog Refreshed

Documentation & install

Readme and setup notes from the catalogue, plus a client-ready config you can copy for your MCP host.

Installation

Add the following to your MCP client configuration file.

Configuration

View docs

{
  "mcpServers": {
    "rightnow-ai-forge-mcp-server": {
      "command": "npx",
      "args": [
        "-y",
        "@rightnow/forge-mcp-server"
      ]
    }
  }
}

Forge MCP Server connects your MCP-compatible AI coding agent to Forge, enabling automated swarm-based optimization of PyTorch code into fast CUDA/Triton kernels on real datacenter GPUs. You submit PyTorch code through your MCP client, Forge optimizes it with 32 parallel swarm agents, benchmarks performance, and returns a drop-in kernel with strong speedups and guaranteed numerical correctness.

How to use

You use Forge MCP Server by configuring your MCP client to connect to the local or remote Forge service, then submitting PyTorch code for optimization or asking Forge to generate an optimized kernel from a natural-language description.

Key workflows you can perform: submit PyTorch code to optimize an existing kernel, or describe an operation to generate a production-ready kernel. Forge runs 32 parallel coder+judge agents, compiles and tests kernels on real GPUs, and returns the best result along with speedup metrics and an iteration history. Every optimization costs 1 credit, and credits are charged only for successful runs (speedup >= 1.1x). You can sign in once via browser-based OAuth, and your session and results are managed by the Forge MCP server.

How to install

Prerequisites: you need a compatible MCP client environment and Node.js/npm installed on your machine. The Forge MCP server is designed to be launched from your MCP client using standard package execution commands.

Install and enable Forge MCP in Claude Code on macOS/Linux:

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

Install and enable Forge MCP in Claude Code on Windows:

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

Install and enable Forge MCP in Claude Desktop (macOS) by configuring the MCP server entry with npx and the package name:

"mcpServers": {
  "forge": {
    "command": "npx",
    "args": ["-y", "@rightnow/forge-mcp-server"]
  }
}

Additional client-specific setup

VS Code / Copilot, Cursor, Windsurf, and OpenCode clients all support Forge MCP Server by configuring the same npx-based command to launch the MCP server integration. The Windows variants use cmd with the /c flag to run the npx command. Use the following pattern in the client configuration: command: npx, args: ["-y", "@rightnow/forge-mcp-server"] (or the equivalent Windows cmd invocation shown in each client guide).

Available tools

forge_auth

Authenticate with the Forge service via browser-based OAuth. Opens a sign-in flow and stores tokens locally.

forge_optimize

Submit PyTorch code to optimize. A 32-agent swarm generates, benchmarks, and returns the best kernel with speedup metrics.

forge_generate

Generate an optimized GPU kernel from a natural-language description. Forge creates a PyTorch baseline and optimizes it into Triton or CUDA.

forge_credits

Check your current Forge credit balance.

forge_status

Check the status of a running or completed optimization job.

forge_cancel

Cancel a running optimization job.

forge_sessions

List past optimization sessions with results.