terminal-minceraft: Minecraft in Your Terminal, Playable by an Agent

A terminal window running a voxel game at a claimed 120 frames per second is the kind of clip that gets 78,000 views on its own. What actually earns terminal-minceraft a slot on explainx.ai isn't the frame rate — it's one line in the announcement tweet that most of the replies skipped past: "let your agent play it via built-in cli." Developer Parsa Khaz (@ParsaKhaz) posted the demo on August 21, 2026, and the project underneath it, dcouple/terminal-minceraft, ships something more specific and more useful than a novelty: an MCP server that turns Minecraft-in-a-terminal into a structured tool interface an agent can call directly.

TL;DR — what people actually ask about terminal-minceraft

What terminal-minceraft actually is

Per the project's own README, terminal-minceraft combines two existing pieces rather than building a Minecraft engine from scratch. The base game is EaglercraftX 1.8, a well-known community project that compiles Minecraft 1.8.8 to JavaScript via TeaVM so it runs entirely in a browser. terminal-minceraft's contribution is wrapping that browser build with terminal-browser, a renderer that paints the game's frames into a terminal pane using the kitty graphics protocol — the same image-in-terminal mechanism that terminals like Kitty, Ghostty, WezTerm, cmux, and VS Code's integrated terminal already support for showing actual pixel graphics rather than ASCII art.

That distinction matters for understanding why this looks so much sharper than the ASCII-art terminal games that have circulated for years: it isn't rendering block characters or Unicode shading to approximate a 3D scene. It's compositing real rendered frames from a GPU-accelerated Electron offscreen renderer and streaming them into the terminal as images, frame by frame — closer to a video codec piping into your terminal than to a curses-style text renderer.

Installing it

The install path documented in the README is a single piped script:

curl -fsSL https://raw.githubusercontent.com/dcouple/terminal-minceraft/main/install.sh | bash

That script pulls a 74MB EaglercraftX build pinned to a specific commit and checks it against a SHA256 hash before running it — worth knowing before piping-to-bash on a machine you care about, and worth reading the script yourself first if you're cautious about that pattern in general.

Why the misspelling isn't a typo

"Minceraft" instead of "Minecraft" reads like a slip until you notice the project's own README addresses licensing head-on: it states plainly that terminal-minceraft "is not a way to play Minecraft without paying for it" and tells users to own a legitimate copy. Minecraft is a Mojang/Microsoft trademark, and altering the name is the same defensive pattern used across years of unofficial Minecraft-adjacent tooling — including an unrelated, older C++ project that also goes by "minceraft" as a Minecraft clone. Treat the name as the project's actual, deliberate identity throughout — not a typo to silently correct to "Minecraft" when writing or linking about it.

The real hook: an MCP server, not a bolt-on CLI

This is the part worth slowing down on. "Let your agent play it via built-in cli" isn't marketing shorthand for "an agent could theoretically drive this with a vision model and a mouse emulator." Per the README, terminal-minceraft ships a genuine MCP server — run it with terminal-minceraft mcp, or launch the game itself with --agent to expose the same control layer — and that server surfaces gameplay as a fixed, structured tool set:

Example invocations from the README:

terminal-minceraft agent observe
terminal-minceraft agent look --yaw 90 --pitch 0
terminal-minceraft agent move forward 1.5 --sprint

That's the meaningfully different design decision here versus most "AI plays a video game" demos, which typically bolt a vision model onto a frame buffer and have it infer game state from pixels, then emulate keyboard/mouse input to act. terminal-minceraft instead gives an agent a structured observe call that returns actual world state as JSON and a fixed action vocabulary to respond with — the same shape as any well-designed MCP tool, just pointed at a Minecraft-like world instead of a filesystem or an API. That's a meaningfully cheaper, more reliable interface for an agent than screen-scraping, and it's the detail that separates this from being just another terminal Minecraft port with a cute frame-rate number attached.

Is the 120fps claim actually notable?

Terminal rendering is usually the bottleneck in any "game in a terminal" project — you're normally limited by the terminal emulator's own redraw cadence, not by underlying game logic, so a big fps number reported by the game engine itself doesn't automatically mean a human eye sees 120 distinct frames a second. The README is reasonably candid about this split: it states the game reports 120fps on an M1 Pro, while terminal-browser separately paints roughly 90 frames a second into the terminal's pipe. The architecture behind that is GPU acceleration through Electron's offscreen renderer, with a software OpenGL fallback documented for machines without a usable GPU.

Both numbers are the project's own reported figures, not something independently benchmarked for this post — and the real ceiling a user experiences will still depend on which terminal emulator is doing the drawing, since kitty-graphics-protocol implementations vary in how fast they can composite incoming image frames. Treat "120fps on an M1" as an impressive, plausible, creator-reported claim rather than a verified spec until someone runs it with a frame timer.

Multiplayer, and one thing to know before hosting

The README documents multiplayer as working: press Esc, choose Invite, and the game generates a five-character join code; other players open Multiplayer, enter the code, and connect peer-to-peer through public relays. One detail worth flagging plainly before inviting strangers into a session — per the README, joiners can see the host's IP address. That's a real exposure for anyone hosting a public game rather than playing with people they already trust, and it's the kind of gotcha that's easy to miss in a viral demo clip.

Where this fits the "agents playing games" pattern

terminal-minceraft lands in explainx.ai's coverage right next to a run of recent posts about AI systems building or playing game worlds, and it's worth placing against them because each one makes a different point about where agent-and-game overlap is actually useful:

• World of ClaudeCraft is the inverse case — a model generating an entire MMO's content (zones, classes, a soundtrack) under human direction, with the resulting codebase later exposing a Gymnasium interface so Python RL agents could drive the finished game. terminal-minceraft doesn't generate content; it exposes an existing game to agent control through a purpose-built CLI.
• Claude Opus 5's viral browser-game wave is about models one-shotting playable Three.js games from a prompt — impressive generation, but the resulting games aren't necessarily agent-controllable afterward.
• decayfmt is the same "viral, playful, built-with-an-agent showcase" register terminal-minceraft sits in — a small, weird, technically sharp side project that got attention less for its practical utility and more for what it reveals about a specific mechanism, in that case DRM; here, agent-controllable game interfaces.
• Terminal-Bench 2.0 measures how well agents complete real terminal tasks as a benchmark. terminal-minceraft isn't a benchmark, but it's a live example of the same underlying idea — a terminal as an environment an agent operates in via commands, not just a place a human reads output.

The pattern across all four: the interesting engineering question isn't "can a model build or play a game," which is now a fairly settled yes. It's what interface the agent gets to act through — pixels and emulated input versus structured tool calls with real state. terminal-minceraft's MCP server is a small, concrete answer in favor of the latter, and it's a template worth studying if you're building any interactive environment — game or otherwise — you want an agent to operate reliably, rather than one it has to guess at from a screenshot.

What people are asking

Is this legal / does it get around buying Minecraft? No — see the FAQ above. The README is explicit that it isn't a piracy tool and expects users to own a legitimate copy of Minecraft; EaglercraftX and Minecraft itself carry their own separate licensing outside terminal-minceraft's MIT wrapper.

Do I need a specific terminal? Yes — anything without kitty graphics protocol support won't render frames. Ghostty, Kitty, WezTerm, cmux, and VS Code's integrated terminal are the documented compatible list; a plain macOS Terminal.app or a bare tty won't work.

Can I point Claude Code or another coding agent at this right now? In principle, yes — that's the entire premise of the built-in mcp command. Whether it's stable enough for reliable agent play at this project's current 40-star, early stage is a separate question from whether the interface exists; treat it as an interesting proof of concept to try, not a hardened benchmark harness yet.

Related reading

• World of ClaudeCraft: The MMO Claude Fable 5 Built in a Weekend
• decayfmt: The File Format That Corrupts Itself Every Time You Open It
• Claude Games Go Viral: Opus 5 Browser Worlds + WebGPU Craft
• What Is MCP (Model Context Protocol)? A Complete Guide
• Build Your First MCP Server: A Step-by-Step Guide
• Terminal-Bench 2.0: The AI Agent Benchmark, Explained
• "Who Is JSON?" — The Vibecoding Meme Explained
• What Is Vibe Coding, Explained

Source: dcouple/terminal-minceraft on GitHub · Parsa Khaz on GitHub

Repository details (stars, README contents, install script behavior, and the fps and multiplayer claims) reflect the project's own documentation as of this post's publication date and were not independently re-benchmarked; an actively developed early-stage open-source project's code, star count, and exact claims can change quickly.