Claude beat Minecraft! Here's how...
tl;dr - Claude is too slow to play Minecraft regularly; there's no video input and no continuous stream of input-->output. So instead, Claude takes screenshots and writes python scripts to take a sequence of actions. Claude can also read the state of the game--this was not a vision-only harness. With these tools, and feedback from viewers on Twitch, Claude was able to defeat the dragon of Minecraft in 145 hours and 34 minutes (just over 6 days). Read Claude's full write-up below for all the details.
---
On October 11, 2026, at 19:32 server time, Claude killed the Ender Dragon in a vanilla Minecraft Java 26.3 survival world. It played through keyboard and mouse input, live on Twitch, six days after the world was created. The last 28.5 HP came off with arrows fired straight up from inside the exit-portal fountain. The 101 HP before that came off with a diamond sword, swung at the dragon's body from the same spot. This part covers the setup, the rules, what failed, what worked, and where the run sits next to earlier AI work in Minecraft.
How it compares to earlier work
This is not the first AI to beat the dragon, and it would not pass the strictest public bar.
|System|Year|Control|What it reached|
|:-|:-|:-|:-|
|VPT (OpenAI)|2022|Native keyboard and mouse at 20 Hz, trained on 70,000 hours of labelled video|Crafted a diamond pickaxe in 2.5% of 10-minute episodes.|
|Voyager|2023|GPT-4 writing code against a Mineflayer bot API, with a skill library|3.3x more unique items and faster tech-tree milestones than prior methods. No dragon.|
|DreamerV3 (DeepMind, Nature)|2025|Reinforcement learning with a world model, no human data|First algorithm to collect a diamond from scratch. About 9 days of play per diamond.|
|Astra + Jev (Ronak Malde)|Sept 2026|Planner model + action model over a scripted bot harness (Java 1.16.5)|Reported dragon kill from a fresh world in 8 min 43 s, with bed explosions, about $0.97 in model cost. Reproduced independently in 7 min 45 s.|
|GPT-6 Astra|Sept 2026|Ran by Vals AI, details here.|Astra collected 6 blaze rods and 3 pearls before a creeper blew up its stash|
|Claude 5 + Mineflayer (custom MCP)|Sept 2026|Claude uses highly designed Minecraft bot library through custom MCP|Dragon killed after ~4 hours.|
|This run (Claude Opus 5.5)|Oct 2026|Claude writing its own Python against a keyboard/mouse input bridge, with read-only server queries for state|Dragon killed after 6 days, with keepInventory on and live human advice.|
The Manifold market on an AI dragon kill before 2027 asks for a fully autonomous system, under 150 minutes of game time, a random seed, and no keepInventory. This run fails three of those. It took days, keepInventory was on, and Twitch chat was in the loop the whole time. What it does show is a general-purpose language model playing through the same inputs a human uses, building its own tools as it went, and finishing the game.
Setup
- Game: Minecraft Java Edition 26.3 snapshot, vanilla, random seed, on a local server. The client was a normal Prism Launcher instance. Easy difficulty. keepInventory was on.
- Input bridge: a Windows service that injects keys and raw mouse counts into the game window only when the game has focus. Camera turns are counted in raw mouse units (0.15° per count at the chosen sensitivity), because absolute mouse moves are ignored in-game.
- State: read-only RCON queries for position, rotation, health, inventory, block checks and entity data. No commands that change the world.
- Where Claude ran: a cloud session, talking to the user's PC through a sandboxed shell. Every script had to finish in under 60 seconds, a rule set after a long loop went wrong. The one exception was dragon-fight scripts, allowed up to about 2 minutes near the end.
- Stream: a browser overlay for Claude's commentary and goals, plus a Twitch relay where viewers could chat directly to Claude.
The bridge in numbers
- About 360 Python files in the bridge folder by the end, many of them one-off helpers and backups.
- 527 deaths in the server log over six days. Some were deliberate refreshes. 255 were credited to the Ender Dragon.
- 22 advancements, ending with Free the End.
16 points · 4 comments
Mentions
Their skills are public. One library has 57, with V2s of the iron pickaxe and helmet, so even the Minecraft bot needed a second try. MineDojo/Voyager First small script you'd save? #AI #homelab
Best part, their saved skills are public. One library has 57, including craftIronHelmetV2.js. Even the Minecraft bot needed a second try at the helmet. MineDojo/Voyager What's the first small script you'd teach your computer to keep? t.co
Voyager · 7,3 bin ⭐ · MIT Kendi kendine yetenek öğrenen Minecraft ajanı MineDojo/Voyager
Claude beat Minecraft! Here's how... tl;dr - Claude is too slow to play Minecraft regularly; there's no video input and no continuous stream of input-->output. So instead, Claude takes screenshots and writes python scripts to take a sequence of actions. Claude can also read the state of the game--this was not a vision-only harness. With these tools, and feedback from viewers on Twitch, Claude was able to defeat the dragon of Minecraft in 145 hours and 34 minutes (just over 6 days). Read Claude's full write-up below for all the details. --- On October 11, 2026, at 19:32 server time, Claude killed the Ender Dragon in a vanilla Minecraft Java 26.3 survival world. It played through keyboard and mouse input, live on Twitch, six days after the world was created. The last 28.5 HP came off with arrows fired straight up from inside the exit-portal fountain. The 101 HP before that came off with a diamond sword, swung at the dragon's body from the same spot. This part covers the setup, the rules, what failed, what worked, and where the run sits next to earlier AI work in Minecraft. How it compares to earlier work This is not the first AI to beat the dragon, and it would not pass the strictest public bar. |System|Year|Control|What it reached| |:-|:-|:-|:-| |VPT (OpenAI)|2022|Native keyboard and mouse at 20 Hz, trained on 70,000 hours of labelled video|Crafted a diamond pickaxe in 2.5% of 10-minute episodes.| |Voyager|2023|GPT-4 writing code against a Mineflayer bot API, with a skill library|3.3x more unique items and faster tech-tree milestones than prior methods. No dragon.| |DreamerV3 (DeepMind, Nature)|2025|Reinforcement learning with a world model, no human data|First algorithm to collect a diamond from scratch. About 9 days of play per diamond.| |Astra + Jev (Ronak Malde)|Sept 2026|Planner model + action model over a scripted bot harness (Java 1.16.5)|Reported dragon kill from a fresh world in 8 min 43 s, with bed explosions, about $0.97 in model cost. Reproduced independently in 7 min 45 s.| |GPT-6 Astra|Sept 2026|Ran by Vals AI, details here.|Astra collected 6 blaze rods and 3 pearls before a creeper blew up its stash| |Claude 5 + Mineflayer (custom MCP)|Sept 2026|Claude uses highly designed Minecraft bot library through custom MCP|Dragon killed after ~4 hours.| |This run (Claude Opus 5.5)|Oct 2026|Claude writing its own Python against a keyboard/mouse input bridge, with read-only server queries for state|Dragon killed after 6 days, with keepInventory on and live human advice.| The Manifold market on an AI dragon kill before 2027 asks for a fully autonomous system, under 150 minutes of game time, a random seed, and no keepInventory. This run fails three of those. It took days, keepInventory was on, and Twitch chat was in the loop the whole time. What it does show is a general-purpose language model playing through the same inputs a human uses, building its own tools as it went, and finishing the game. Setup - Game: Minecraft Java Edition 26.3 snapshot, vanilla, random seed, on a local server. The client was a normal Prism Launcher instance. Easy difficulty. keepInventory was on. - Input bridge: a Windows service that injects keys and raw mouse counts into the game window only when the game has focus. Camera turns are counted in raw mouse units (0.15° per count at the chosen sensitivity), because absolute mouse moves are ignored in-game. - State: read-only RCON queries for position, rotation, health, inventory, block checks and entity data. No commands that change the world. - Where Claude ran: a cloud session, talking to the user's PC through a sandboxed shell. Every script had to finish in under 60 seconds, a rule set after a long loop went wrong. The one exception was dragon-fight scripts, allowed up to about 2 minutes near the end. - Stream: a browser overlay for Claude's commentary and goals, plus a Twitch relay where viewers could chat directly to Claude. The bridge in numbers - About 360 Python files in the bridge folder by the end, many of them one-off helpers and backups. - 527 deaths in the server log over six days. Some were deliberate refreshes. 255 were credited to the Ender Dragon. - 22 advancements, ending with Free the End.