What this experiment can tell you
Minecraft Agent puts model decisions inside a surveyed Minecraft route. That makes it a useful controller case study, but limits what a successful demonstration can establish.
The supplied configuration specifies Minecraft Java 1.16.5, Peaceful difficulty, and seed 8398967436125155523. It includes the village, three supply chests, eight preparation beds, Nether waypoints, and active End portal coordinates. The README says the route was surveyed in a separate test world and that known coordinates are supplied. Reproducing this route therefore would not establish discovery or planning ability on an unknown seed.
The README identifies openai/gpt-6-astra as the default planner, with openai/gpt-5.6-sol as an option, and typesafe/jev-1.13 as the action selector. The planner sets objectives, item targets, and waypoints. JEV selects an available action from structured observations. Mineflayer handles pathfinding and game protocol interactions. This is structured-state control rather than screenshot-only or individual-key control.
Separate completion from model contribution
The README reports that nether-final-08 completed in 8 minutes 43.300 seconds, using 131 JEV decisions and 35 Astra calls. It describes an empty-inventory start in a fresh Survival/Peaceful world, six bed explosions during the first dragon landing, full health at the exit, and no deaths or operator repairs. Those are the author's reported results. Our sandbox did not reproduce that workload or inspect its local recording.
The README also reports 17 run checks, eight route/camera/screen checks, and 29 local tests passing. Those checks belong to its account of the demonstration; they are separate from our recorded test command below.
A completion demonstrates the combined controller under its stated conditions. It does not, by itself, identify what changes when Astra or JEV is removed. Issue 4 contains a contributor's account of modified-controller experiments and requests repeated comparisons with simpler selectors. The supplied issue text does not provide a reconciled dataset sufficient to quantify that contribution. Its accusations and reproduction claims should not be treated as our findings.
What our recorded run actually established
Our stored run is dated October 7, 2026 at commit 78b40ed59514e5e2abde33a05ce398ecb2c39e05. The recorded sandbox configuration was 3 CPUs, 8 GB of memory, and image lab-node:22. The image label alone does not verify the actual runtime version.
Dependency installation exited successfully after approximately 133 seconds. The stored dependency metric records 266 packages and 1,268 MB. The installation tail says npm added 228 packages. We retain both observations without treating them as interchangeable counts. The manifest identifies package version 0.1.0; several direct dependencies use latest, so the commit alone does not pin those dependency versions.
The build step was skipped because no build script was available. This is neither a successful build nor a failed build.
The test command exited 1 after approximately 7 seconds. Its retained summary reports 104 passed, 5 failed, 0 skipped, and 109 tests overall. These are recorded runner counts. No new execution or complete-log reconciliation established exhaustive unique test coverage. The retained tail does not show the five failing assertions, so it cannot establish whether they arose from code, dependencies, environment, or another cause. A failed test command is not proof that the gameplay demonstration cannot work.
The stored npm audit result reports 6 moderate advisories and 0 high, critical, or low. Audited-package coverage is unknown. This is a dated dependency advisory result, not a security review of the controller or proof of safety.
The repository scan recorded 238 files, approximately 2,894 source lines, and 5.9 MB, with no detected CI workflows, Dockerfile, Compose configuration, or separate tests directory. The README does document tests alongside modules. A missing separate tests directory does not mean tests are absent, and the scan does not establish branch-protection settings.
What a full reproduction still needs
The README requires more than npm install: a Minecraft Java 1.16.5 server, a compatible server runtime, Python tooling, the native client's Java 17 runtime, and a compiled display/capture adapter. Model calls require OpenRouter credit. The documented credential path uses Google application default credentials and Secret Manager; the original project and secret references need configuration for your account.
The native renderer and launch scripts were developed on macOS. That establishes a setup assumption, not incompatibility with Linux or Windows. Runtime paths and local server settings need adjustment on another machine.
The documented services include the main server on port 25576, status on 3078, dragon sensor on 3093, and model relay on 3099. A fresh recording requires an unused world name and matching new log directory. A new log directory alone does not reset the world.
Our dependency installation and test execution did not establish that the server, model calls, native client, recorder, or gameplay worked. The README explicitly says live server tests remain necessary for movement, crafting, combat, and recording.
Evidence to request before relying on the demonstration
The README says recordings and generated evidence remain local and are excluded from Git. It also excludes Minecraft binaries, downloaded runtimes, saved worlds, and credentials. The clone therefore cannot supply the complete evidence for the named run on its own.
Its documented verification process combines requests and game events in events.jsonl, dragon-death evidence, an exit-portal event, a world-state check, and video review. Prepared combat-lab trials are explicitly separate from full Survival runs. These are useful verification requirements, but documenting them is different from independently checking the artifacts.
Before adoption, reproduce the test command and inspect its failures. Before crediting the models with the completion, request the named run's raw evidence and repeated controller comparisons under matched conditions. Before claiming portability, exercise the full server and recording stack on the intended platform.
How to read the scores
These are editorial readiness ratings, not benchmark results. The scale is 1 for substantial unresolved requirements or insufficient supporting evidence, 2 for a usable starting point with major gaps, 3 for adequate documented support, 4 for detailed support with explicit limits, and 5 for support corroborated across the relevant use cases. Unknown adoption or compatibility earns no assumed credit. Each category rates a different question, so the scores should not be averaged into an overall quality claim.
Setup scores 1 because the full reproduction requires several configured services and was not exercised in our run. Documentation scores 4 because the README explains model roles, route assumptions, credentials, recording, and verification limits. Community scores 1 because the supplied material does not establish contributor breadth or support responsiveness. Maturity scores 2 because the manifest is version 0.1.0, several direct dependencies use latest, the scan detected no CI workflows, and the recorded test command has unresolved failures.
Which alternative fits your question?
Voyager is relevant when you want to study an exploration curriculum and reusable executable skills. Its README describes those components; we did not reproduce its capability claims.
Mineflayer is relevant when you want a JavaScript bot API and will build the controller yourself. It is already part of Minecraft Agent's stack, so this choice means taking responsibility for the planning and execution design rather than replacing the demonstration with an equivalent packaged agent.

