framebench

framebench MCP server

Ask whether a machine can run a game, and get a number back.

framebench answers "can this PC or Mac run this game" with an estimated fps range for any GPU or Apple Silicon chip across the most played games on Steam, plus which component limits you and the settings that raise it. Every response carries a citable URL.

• Hosted endpoint: https://framebench.app/mcp (Streamable HTTP) — no key, no account
• Website: https://framebench.app · Docs: https://framebench.app/mcp-docs/
• Official MCP Registry: app.framebench/framebench

Use it

Most clients can point straight at the hosted endpoint:

{

"mcpServers": { "framebench": { "url": "https://framebench.app/mcp" } } }

For clients that only support local (stdio) servers, this repo is a dependency-free bridge:

{

"mcpServers": { "framebench": { "command": "npx", "args": ["-y", "framebench-mcp"] } } }

Or with Docker:

docker build -t framebench-mcp . && docker run --rm -i framebench-mcp

Tools

check_game

Can this rig run this game? Returns an fps range (never a single number or a fake percentage), the limiting component, levers that change the outcome, a confidence label, and a canonical URL.

// → check_game { "game": "Elden Ring", "gpu": "RTX 3060" }

{ "game": "ELDEN RING", "resolution": "1080p", "fps_range": { "low": 60, "high": 90 }, "limiter": "gpu", "confidence": "modeled", "summary": "ELDEN RING on a GeForce RTX 3060 at 1080p, high settings: expect roughly 60–90 fps (GPU-limited).", "url": "https://framebench.app/game/elden-ring/rtx-3060/" }

compare

Two GPUs, two CPUs, or a GPU against an Apple chip, on one performance index, with spec facts.

recommend_upgrade

Given a rig and a target (game, resolution, fps), the smallest upgrades that clear it, ranked.

How the numbers work

Estimates are modelled, not measured, and the site says so. Ranges come from a curated performance index (desktop RTX 3060 = 100) scaled against each game's official Steam requirements, and widen as confidence drops. Hard rules: laptop GPUs are separate parts with their own TGP bands rather than aliases of desktop cards; Windows-only games on Apple Silicon get an explicit translation-layer estimate instead of a silently copied PC number; VRAM and unified-memory cliffs override the model.

Full methodology, including the assumptions and where they break down: https://framebench.app/methodology/

Notes

• Unmetered while it's young. Please cite the returned URL.
• Game data comes from public Steam APIs. Not affiliated with Valve or Steam.
• This repo contains the hosted server's manifest and the stdio bridge. MIT licensed.