Give an LLM a body.

A Minecraft bot that is a real agent — it perceives, acts, remembers, survives, hires help, and answers you by chat, voice or phone.

git clone https://github.com/cagataycali/strands-minecraft

Read the README MIT · Node ≥ 22 · Strands + mineflayer

First-person view from the bot: a sandy beach, two turtles at the waterline, a yellow sponge build on the right — Every frame on this page is the bot's own first-person view, recorded off its live dashboard — this one is 4 s of `demo-creeper.gif`, uncut.

60: tools
100%: of mineflayer covered
74: test files
18.6k: lines of TypeScript

Watch it think.

Two uncut captures from the bot's eyes. One instructed, one not.

Animation: the bot digs the blocks in front of it, then places them back — `you>` dig the blocks in front of you, then put them back

Animation: a creeper approaches on a beach and the bot backs away, uninstructed — creeper encounter, uninstructed

Try it. One line, end to end.

One real exchange from the README, in the grammar src/index.ts prints: prompt, tool call, reply, status block, journal.

You say

you> keep mining until you have 64 iron, I'll be around

It calls a tool

# → start_journey goal="mine until 64 iron_ingot in inventory"

It answers in the game

🤖 On it — journey jm3kq2 started. Ask me anything while I dig.

Every step opens with the digest — formatDigest, ≤ 600 chars, zero tool calls

Your status right now:
HP 20/20 food 17/20 · (118, 42, -38) overworld, survival
time: day
threats: none in radar range
bag: 4 meal(s), 12 torch(es), 41 building block(s) | tools: iron_pickaxe, stone_shovel | weapons: iron_sword
journey: jm3kq2 "mine until 64 iron_ingot in inventory" step 1 — last: find_blocks iron_ore,deepslate_iron_ore exposed=false → 3 veins; tunnelling

It works while you don't — each line ends in a measured [Δ …], not the model's account of itself

🧭 [jm3kq2 #1] find_blocks iron_ore,deepslate_iron_ore exposed=false → 3 veins; tunnelling  [Δ +9 cobblestone]
🧭 [jm3kq2 #2] dig_vein at (118, 42, -37) → 5 raw_iron  [Δ +5 raw_iron, moved 21m]

Same lines as the README's Ten minutes with it; the digest is the output of src/digest.ts for that step.

One session. Seven rails.

Game chat, the you> prompt, push-to-talk, a realtime call, the dashboard, a phone call, your tiny — all write into one shared history.

Send another request while one runs: the session forks the history, both turns think, the finished one folds back in completion order.

A nervous system, not a chat loop.

Most of what keeps the bot alive never touches the model.

⚡ Reflexes

300 ms tick · zero tokens

Lava → water bucket or flee. About to die → disengage. Creeper in blast radius → sprint. Stuck 20 s → shake loose. Idle: eat, wear better armor, grab drops.
👂 Sentinel

event-driven · zero tokens

Hears primed TNT and creepers, radar-tracks hostiles, names your attacker, notices dusk, chests opened behind your back — and tells the agent in one edge-triggered note.
📋 Digest

per self-prompt · ≤ 600 chars

Every thinker cycle and journey step opens with one status block — vitals, threats, bag, journey, workers — so the model reasons from state instead of spending a tool call asking.
🧭 Critic & ledgers

per journey step

Each step's journal line gets a measured [Δ +12 cobblestone, -2 hp, moved 34m] tail — narration can't hide a stall. Finished goals land in a completed ledger, failures in too_hard.

Say it once. Walk away.

Journeys iterate toward a goal for hours. The fleet hires worker bots — real second bodies — when a job splits.

strands-bot — tsx src/index.ts

⛏️ strands-minecraft — connecting… 🧠 model bedrock · global.anthropic.claude-sonnet-5 · us-west-2 ✅ Spawned as StrandsBot at (12.5, 64, -87.5) you> keep mining until you have 64 iron 🧭 [jm1x9k2a #1] Found an exposed iron vein 14 blocks north; dug it out. [Δ +6 raw_iron, moved 21m] 👥 [Chopper #1] joined the server, starting: "fell the oaks by the water, 32 logs" 💭 Dusk. The player is 40 m off and safe; I'll eat before the next step.

60 tools. If mineflayer can do it, the agent can.

Every game mechanic has a tool or a written reason it doesn't — the audit.

Hover or tap a tool — the line is its real description: from the source, first sentence, nothing rewritten.

Your bot as a device.

Optional. Enroll it and your personal AI at tiny.technology sees through its eyes and sends it to work — from web, phone or watch.

enroll

npm run enroll -- --endpoint https://minecraft.yourdomain.com
# = npx tiny-tech enroll --body strands-the-miner · or set MINECRAFT_PUBLIC_URL

The endpoint the enrollment reads — one public route, six behind a bearer token, 5 req/s per token
route	gate
`GET /api/health`	public
`GET /api/telemetry`	token
`GET /api/camera/snapshot`	token
`GET /api/stream.mjpeg`	token
`GET /api/events`	token
`POST /api/chat`	token
`POST /api/stop`	token

Run it.

The whole rig — Minecraft server, bot, dashboard — in one command. Or bare metal, if you want the microphone.

Docker

git clone https://github.com/cagataycali/strands-minecraft && cd strands-minecraft
cp .env.example .env        # AWS_* for Bedrock by default
docker compose up -d        # server + bot · dashboard on :3008

Bare metal · Node ≥ 22

npm install
cp .env.example .env        # credentials + MC_HOST/MC_PORT
npm start                   # bot joins, you> prompt opens

Bedrock by default; OpenAI or Anthropic with one env var (STRANDS_MODEL_PROVIDER) and their SDK package. Give the Docker VM at least 6 GB (8 GB with fleet workers) — the bot checks its headroom at boot and says so in one line.

How it compares.

Voyager and Mindcraft are about learning to play. This is about a body. Every claim below is quoted from each project's own README.

criterion	strands-minecraft	Mindcraft	Voyager	mineflayer alone
the model drives	60 typed tools = 100% of mineflayer	chat actions; optionally "write/execute code on your computer" (off by default)	"an ever-growing skill library of executable code"	your JavaScript/Python
models	Bedrock default; OpenAI, Anthropic via one env var	18 APIs incl. openai · google · anthropic · ollama (local)	"OpenAI's GPT-4"	none
when the model is asleep	300 ms reflexes + sentinel, zero tokens	—	—	everything is your code
more than one body	`manage_bots` hires workers; independent compose projects	`--profiles andy.json jill.json`	a single agent	as many as you write
how you reach it	game chat · CLI · push-to-talk · realtime call · dashboard · phone · tiny	game chat · web UI (`:8080`) · TTS narration (`speak_model`)	Python API (`voyager.learn()`)	your code
runs on	Node ≥ 22, Docker compose with the server	Node 18/20, Minecraft ≤ 1.21.11	Python ≥ 3.9 + Node ≥ 16, Fabric 1.19 mods	Node, Minecraft 1.8 → 26.1

Same rows as the README's How it compares.

Read deeper.

The audits and the incident write-ups that shaped the code. Each hook is the file's own first sentence; reading time is counted, not guessed.