We were wrong about modded worlds
This site used to say our AI could not join a heavily modded server. We had not tried it. Somebody asked whether we had, so we went and did, and the claim was wrong. This is what actually happens, measured on a 49-mod pack.
The assumption was reasonable and completely untested: a big modpack changes so much about a world that a bot built against vanilla data would not survive contact with it. It sounds right. It is the sort of thing that gets written into a FAQ because nobody wants to promise something they have not checked.
But an untested caution is still a claim. We spun up the Cobblemon Fabric pack, 49 mods on Minecraft 1.21.1, and pointed the bot at it.
-
It joined, and nothing broke
A 49-mod Cobblemon pack on Fabric, 1.21.1. The bot connected as an ordinary player, stayed connected, walked 8.2 blocks under its own power, and reported 20/20 health with zero packet parse errors during play. Everything we had assumed about protocol incompatibility was wrong. Mods overwhelmingly add content rather than protocol, and content travels over the same packets vanilla already uses.
-
But it could not name anything
mineflayer resolves blocks through minecraft-data, a static file written per Minecraft version before any of those mods existed. So a Cobblemon apricorn log arrived with the correct network state id, 43707, and an empty string for a name. The bot could see it perfectly and had no idea what it was.
-
The server already knows, and will tell you
A server has to know its own blocks or it could not place one, and tab completion is how it tells you. Type "/setblock 0 64 0 " and it answers with every block it will accept. Walking that one prefix at a time returned 1,367 blocks, 304 of which vanilla has never heard of, plus 2,039 items and 135 entities. It took under a second and needed no support for any particular mod.
-
Naming a block is not being able to touch it
Even with names, mineflayer had no hardness and no harvest rule for those blocks, so bot.dig() returned successfully and broke nothing. Success, reported, with the block still standing. The fix is calibration: place one of each on a scratch block, write down the state id the bot reports, and put a synthetic entry in the registry that borrows a lookalike vanilla block’s weight. 304 blocks in 140 seconds, once per server.
Two traps in the calibration
Neither announced itself, and both would have produced a registry that looked complete and was quietly wrong.
A failed setblock leaves the previous block standing.
The read then records its state id under the new block's name,
overwriting the correct entry made for it a moment earlier. One failure
corrupts two blocks and reports nothing. Clearing to air first turns
that into a reading of air, which is countable.
A block with several states answers differently each time. Think of a berry that grows, or a sign that faces somewhere. One reading names one state and silently leaves the rest unknown. Placing twice and keeping only the blocks that agree sets those aside instead.
What it can do now, and what it still cannot
- blocks the server offered
- 1,367
- that vanilla has never heard of
- 304
- to learn all 304
- 140s
- mined, ore included
- 8/8
- items learned in 30 s
- 705
The ore is worth a note. It resisted a first attempt and we nearly wrote it up as a limitation. It was resisting a bare hand. With a pickaxe it broke. That is Minecraft working correctly, and a good reminder that a failed measurement is a claim about your setup before it is a claim about the world.
Crafting is the honest remaining edge. A mod that adds its own furnace, its own recipes and its own machine to run them is a mod Cog can build with and mine in, and cannot yet craft in. Knowing block names does nothing for that, so we are not claiming it.
We had a sentence on a public page that was wrong because checking it looked like work. Checking took an afternoon. If you read a claim on this site and wonder whether anybody measured it, the answer should be in a post like this one.