TokenAssemble
Can I Run ItGPUsSystemsModelsCheck your machine

Use TokenAssemble inside ChatGPT & Claude

Connect our free, read-only MCP server and your AI assistant can check whether any machine runs any local model — fit verdict, grade, honest speed tier, VRAM breakdown, assumptions stated. Computed by the same engine as the site, never made up.

Server endpoint

https://tokenassemble.com/api/mcp

Once connected, just ask

“Can an RTX 4090 run Llama 3.1 70B?”
“Will Qwen2.5-32B fit on my 7900 XTX with llama.cpp?”
“What can a Mac Studio M3 Ultra run at 16K context?”
“Can a DGX Spark run gpt-oss-120b, and how fast?”

How to connect

Claude (web — easiest)

  1. Open Settings → Connectors in Claude.
  2. Click Add custom connector.
  3. Paste the endpoint URL and save. No extra software needed.

Claude Desktop

Open Settings → Developer → Edit Config, add the block below to claude_desktop_config.json, then fully quit Claude (⌘Q) and reopen.

{
  "mcpServers": {
    "tokenassemble": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://tokenassemble.com/api/mcp"]
    }
  }
}

macOS not finding it? Finder-launched apps get a minimal PATH that misses Node, so "npx" fails silently. Use this variant — the absolute npx path (yours may differ; run which npx in a terminal) plus an explicit PATH:

{
  "mcpServers": {
    "tokenassemble": {
      "command": "/opt/homebrew/bin/npx",
      "args": ["-y", "mcp-remote", "https://tokenassemble.com/api/mcp"],
      "env": { "PATH": "/opt/homebrew/bin:/usr/local/bin:/usr/bin:/bin" }
    }
  }
}

ChatGPT

Connector support in ChatGPT depends on your plan and the ongoing rollout. If available:

  1. Enable Developer mode / Connectors in ChatGPT settings.
  2. Add a new connector and paste the endpoint URL.
  3. Ask it whether your machine can run a model.

For developers

A remote MCP server over Streamable HTTP — JSON-RPC 2.0 over POST (initialize, tools/list, tools/call). Read-only, no auth, rate-limited per IP. Opening the endpoint in a browser returns an error by design — it's a JSON-RPC endpoint, not a page.

can_i_runFit verdict + grade + speed tier for hardware × model

hardware, model, quant?, context?

Accepts common names (“4090”, “Llama 3.1 70B”) and resolves them to canonical slugs; unknown names return valid suggestions, never invented verdicts. Full input schema comes from tools/list.

Common questions

What is the TokenAssemble MCP server?
A remote Model Context Protocol (MCP) server — the standard way for AI assistants like ChatGPT and Claude to call live tools. Connecting it lets your assistant answer 'can this machine run this model?' with TokenAssemble's computed verdict: fit, grade, speed tier, VRAM breakdown, and every assumption stated. It is free and read-only.
Does the AI make up the numbers?
No. Every answer is computed by the same engine that powers tokenassemble.com, from measured file sizes and published specs. If you name hardware or a model we don't track, the tool says so and suggests valid options instead of inventing a verdict.
Which assistants does it work with?
Any MCP-capable client: Claude (web and desktop), ChatGPT in developer/connector mode, Cursor, and others. The same endpoint works everywhere: https://tokenassemble.com/api/mcp.
It isn't showing up in Claude Desktop — what do I do?
On macOS, apps launched from Finder get a minimal PATH that can miss your Node install, so npx fails silently. Use the absolute path to npx (run 'which npx' in a terminal) as the command, add an env block with your PATH, then fully quit Claude (Cmd-Q) and reopen.
Try it on the site →How the math works