aibengineering/beat-the-game-minecraft

Mine AI MCP — the run that beat Minecraft

An LLM agent played Minecraft 1.21.4 from an empty world to a defeated Ender Dragon, autonomously, in a single unbroken session. No human input after the prompt, no scripted behaviour trees, no save-scumming. This dataset is the complete record of that run.

📺 Watch the run: https://www.youtube.com/watch?v=ZjtwWEfFVFY 💻 Code: https://github.com/aibengineering/mine-ai-mcp (MIT)

What it cost

Beating Minecraft took $93.17 of compute at API-equivalent rates. Where that goes:

Component	Tokens	Rate ($/MTok)	Cost	Share
Cache read	163,325,079	$0.50	$81.66	87.7%
Cache write (1 h TTL)	779,047	$10.00	$7.79	8.4%
Output (incl. 67,071 thinking)	148,383	$25.00	$3.71	4.0%
Input (uncached)	770	$5.00	$0.004	0.0%
claude-opus-5[1m] subtotal			$93.166
claude-haiku-4-5 (1,616 in / 15 out)		$1.00 / $5.00	$0.002
Total			$93.168

No money was actually spent at these rates. This run was played on a Claude subscription, not the API — the system/init record in transcript.jsonl shows apiKeySource: "none", and the session drew on subscription rate-limit windows rather than API billing. The figure above is what the same token counts would cost at first-party API rates, which makes it a useful measure of the work done, not an invoice. A subscription plan does not charge per token.

More on where the cost sits, and what it implies for long-horizon agents, in Notes on cost below.

The run

Outcome	Ender Dragon defeated
Date	2026-09-15, 00:43:30 → 04:40:21 UTC
Model	claude-opus-5[1m], effort high
Seed	97996358
Bot	AI_M1KE
Minecraft	1.21.4 vanilla (server.jar SHA-1 4707d00eb834b446575d89a61a11b5d548d8c001)
Tool calls	388 across 37 distinct MCP tools

Proof of completion, from minecraft-server.log:

[04:39:50] [Server thread/INFO]: AI_M1KE has made the advancement [Free the End]

The agent's own closing summary, in result.json, cites three independent confirmations: attack_dragon_perch reporting dragon_died with health 34 → 0, the ender_dragon entity gone from loaded mobs, and the dragon egg spawned at (0, 61, 0) with the exit End portal open — both of which only generate on a confirmed kill.

The agent in its own words

The agent kept its own notes as it played — five entries in the notes table of bot-data.sqlite, written to survive death or a session restart. The last one was saved at (4, 57, -2) in the End, twenty seconds after the kill:

ENDER DRAGON DEFEATED at 2026-09-15 04:39 UTC. Killed via perch melee with diamond sword after destroying all 10 end crystals (8 by bow, 2 caged by end-stone staircase melee). Final blow at dragon health 34 -> 0.

Context: Completing the playthrough goal. Full run: fresh spawn -> diamond gear -> Nether blaze rods -> 12 overworld ender pearls -> stronghold at -269,-36,-1819 -> End.

The earlier four record a gear milestone, the Nether portal coordinates, the stronghold found by eye-of-ender triangulation "after a 1700-block journey", and one death: AI_M1KE was blown up by Creeper at (-110.5, 32.2, 194.6) — after which the agent noted it had "recovered everything".

Every one of the 387 rows in action_requests also carries a rationale the model wrote before acting, which is where the moment-to-moment reasoning lives:

(locate_stronghold) Travel to the triangulated stronghold; a carved pumpkin is not craftable with these actions and 27 arrows will have to suffice.

(view_blocks) The temporary furnace was left in the world after the smelt was interrupted; I need its coordinates to reuse it for mutton.

Files

File	Size	Contents
transcript.jsonl	4.7 MB	Start here. 1,121 rows: 731 assistant turns, 388 tool results, system/init, final result.
stream.full.jsonl	8.4 MB	13,930 rows — the raw stream including incremental SSE deltas and token-count events.
bot-data.sqlite	20.4 MB	The runtime's own structured record. 19 tables.
minecraft-server.log	5.6 KB	Vanilla server log, including the advancement above.
minecraft-console.log	14 KB	Full server console output.
prompt.txt	3.3 KB	The seed prompt, verbatim — the only instruction the agent was given.
result.json	5.2 KB	Final result object, extracted for convenience.
run.json	3.9 KB	Launch inputs: seed, model, effort, server jar digest, and the full prompt.
world.zip	62.5 MiB	Final world save: Overworld, Nether, End, and saved player data. See WORLD.md.
WORLD.md	—	Restoration instructions and verification details.
world-files.sha256	—	Full SHA-256 for each archived world file.
SHA256SUMS	—	Full SHA-256 for publication files (excluding itself).
MANIFEST.md	—	Artifact inventory and checksums.

transcript.jsonl schema

One JSON object per line. The shape depends on type:

• assistant (731) — a model turn. message.content[] holds thinking (231), text (112), and tool_use (388) blocks.
• user (388) — a tool_result block: what the Minecraft world returned.
• system (1) — subtype: "init": model, tool list, MCP server status.
• result (1) — the final summary, usage, and cost.

The seed prompt is not in the transcript. Every one of the 388 user records is a tool result; there is no initial user message. Claude Code received the prompt as a command-line argument (claude --print "$PROMPT") and --output-format stream-json emits only the output stream, so the instruction reached the model as the first user turn but was never echoed to stdout. It is preserved separately in prompt.txt (and in run.json). Nothing was removed during publication — the stream never contained it, and no synthetic message has been inserted to paper over the gap.

Rows are heterogeneous, so a strict-schema loader may struggle; reading it as raw JSON lines is the reliable path:

import json
rows = [json.loads(l) for l in open("transcript.jsonl", encoding="utf-8")]
tool_calls = [b for r in rows if r["type"] == "assistant"
              for b in r["message"]["content"] if b["type"] == "tool_use"]

bot-data.sqlite

The MCP server's own view of the run, independent of the model's account of it:

Table	Rows
action_requests / action_responses	387 / 387	Every tool call and its outcome
action_executions	239	Execution records with timings
events	320	World events observed by the bot
frontier_chunks	6,836	Explored-chunk index
frontier_chunk_materials	67,878	Materials seen per chunk
frontier_chunk_biomes	10,058	Biomes per chunk
bot_inventory	31	Inventory snapshots
action_incidents	22	Incident references
stronghold_throws	3	Eye-of-ender triangulation
notes	5	The agent's own persistent notes

Comparing action_responses against the transcript is the interesting bit: the model's narration and the runtime's record are separate sources for the same events.

Reproducing

The winning commit is tagged. The release image clones that exact commit and reads nothing from your working tree:

git clone https://github.com/aibengineering/mine-ai-mcp.git && cd mine-ai-mcp
docker compose -f compose.yaml -f compose.beatthegame.yaml build

Commit de4fef3260424d973bf81d0eaa15ff1a2f73ecff, tag beat-the-game. Every forked dependency is tagged mine-ai-beat-the-game in its own repository, so the whole dependency graph is addressable by one name.

A rerun will not replay this trajectory. The seed and settings reproduce the world and the configuration; the agent's play is not deterministic.

Not included

• Minecraft server JAR and libraries — Mojang's, not ours to redistribute. run.json records the official SHA-1 and download URL.
• Incident captures — 63 MB of per-tick combat telemetry across 11 incidents (7 × combat_waiting, 1 × death, 1 × non_entity_damage, 1 × execution_slice_exhausted, 1 × runtime_closed). Available on request.
• Claude Code's native session file — carries account context, so it is withheld. transcript.jsonl and stream.full.jsonl contain the same agent trajectory.

Notes on cost

The $93.17 above is an API-equivalent figure, not a bill. The run used a Claude subscription — transcript.jsonl's system/init record shows apiKeySource: "none" — which does not charge per token. The number is what these token counts would cost at published API rates, included because it fairly measures the compute consumed.

Rates used: claude-opus-5 at $5.00/MTok input, $25.00/MTok output, cache writes at 1.25× input (5-minute TTL) or 2× (1-hour), cache reads at 0.1× input. The 1M context window carries no long-context premium — these are the standard rates.

Three things the arithmetic shows:

• It reconciles exactly. The breakdown reproduces the recorded total_cost_usd to six decimal places ($93.166434 computed vs $93.166435 recorded), and Haiku's $0.001691 matches independently.
• The cache TTL is inferable from the total. At a 5-minute TTL the same token counts come to $90.25; only the 1-hour rate reproduces the recorded figure.
• Caching saved about 89%. Paying full input rate for all 164 M cached tokens would have been $820.52; it came to $89.45 instead.

Even after that saving, cache reads are 87.7% of the total. A long-horizon agent re-reads its entire accumulated context on every turn, so the input side dominates however cheap each read becomes — the ratio of cached to uncached input tokens here is roughly 212,000:1, and output is only 4% of the cost.

Time tells a similar story: of 3 h 56 min wall clock across 389 turns, only 17.8 min (7.5%) was spent inside API calls. The other 92% was the agent waiting on the world — mining, walking, and fighting happen in Minecraft's time, not the model's.

Notes and caveats

• The player name aibengineeringl appears in the logs and database: that is the project owner, present as an observer. No other personal data is included.
• The run used a 1M-context model. Behaviour with smaller context windows will differ substantially on a task this long.
• duration_api_ms counts time inside API calls only; the rest is the world ticking.

Licence

Data: CC BY 4.0. Software: MIT, in the repository above. Minecraft is a trademark of Mojang Studios; this project is unaffiliated with and not endorsed by Mojang or Microsoft.

Citation

@misc{mineaimcp2026,
  title  = {Mine AI MCP: beating Minecraft with an LLM agent},
  author = {AI Bengineering},
  year   = {2026},
  url    = {https://github.com/aibengineering/mine-ai-mcp},
  note   = {Dataset: run that defeated the Ender Dragon, 2026-09-15}
}