@agntn/puzzles

Public crypto puzzles and bounties, as typed records. You ask for a puzzle, you get its address, its key material and what happened on chain.

Why?

Every puzzle thread has the same three things: the addresses, the prizes, and who solved what. Every scanner and tracker re-types them from the thread, slightly differently each time. So they live here once, as code. A puzzle is a TypeScript record. The type checker reads it before a test does. The CLI, the MCP server and the Pi and OMP extensions read one registry.

Docs, one page per puzzle and a live playground: puzzles.agntn.dev.

[!WARNING] Pre-1.0. The API, the CLI flags and the data model can still move. Pin an exact version if you build on it.

✨ Features

• 🧾 Data as code. One PuzzleSpec literal per puzzle, built by one puzzle() factory for every chain. No JSON, no build step.
• 🕳️ Absent means absent. A puzzle without a solver or a prize has no such key. Nothing serializes as null.
• 🔑 Key material in every shape. Hex, WIF, a BIP38 payload, a seed phrase, secret shares, a sealed Arweave keyfile, or just a bit width. Every key of a multisig, too. One builder.
• 💤 Lazy registry. Importing the package loads no records. get("bits/71") imports one collection module.
• ✅ Verification is a value. A published key derives the address or it doesn't. Nothing throws for a bad record.
• 💰 Live balances. puzzle.balance() through @agntn/explorers. Base units as bigint, API keys redacted from errors.
• 👀 A watch on the record. puzzles watch lists the deposits and spends a record misses, a prize that moved, a public key a spend gave away, a source page that changed. It never edits a record. You do.
• 📋 A checklist before the weekend. puzzles eligibility gathers the source, the address, lifetime totals from the explorer, the status with its evidence and what counts as a solution. Whatever nobody can fill comes back as a missing row, never a guess.
• 🖼️ The puzzle itself, not a link to it. puzzles_assets hands a model the image, the stage files and the archived posts behind the hints, checked against the SHA-256 the record pins. Links rot. These don't.
• 🤖 Fifteen agent tools. One executor behind MCP, Pi and OMP. Same answer everywhere.
• 🌐 Runs anywhere. Neutral ESM on the Fetch API. Node, browsers, edge workers.

📦 Install

pnpm add @agntn/puzzles

Node.js 26 or newer for the CLI.

🚀 First call

npx @agntn/puzzles stats

That prints the total, one count per status and how many puzzles have a known public key. No key, no config, no network. The records ship inside the package. The bare puzzles below is pnpm exec puzzles after a local pnpm add, or just puzzles after pnpm add -g @agntn/puzzles.

puzzles show hash-collision/sha256

hash-collision/sha256	unsolved	0.27734251 BTC	35Snmmy3uhaer2gTboc81ayCip4m9DT4ko
chain: bitcoin  address kind: p2sh
hash160: 292fb39df7cd619a396069383928e6bfb74ebec5
redeem script: 6e879169a87ca887 (hash 292fb39df7cd619a396069383928e6bfb74ebec5)
public key: unknown
private key: unknown
started: 2013-09-13 05:59:09
transactions: 1
	funding	2013-09-13 05:59:09	0.1 BTC	397f12ee15f8a3d2ab25c0f6bb7d3c64d2038ca056af10dd8251b98ae0f076b0
explorer: https://blockstream.info/address/35Snmmy3uhaer2gTboc81ayCip4m9DT4ko
source: https://bitcointalk.org/index.php?topic=293382.0

Commands

Command	What it prints
puzzles stats	Totals, status counts, prize sums and the data version
puzzles collections	One row per collection: key, counts, author
puzzles authors [key]	One row per author, or one author's record with its sourced facts
puzzles author <key>	One author's record, straight from the puzzles_author tool
puzzles solvers [key]	One row per named solver, or one solver's record: every solve, profiles and sourced facts
puzzles solver <key>	One solver's record, straight from the puzzles_solver tool
puzzles show <id>	One puzzle's record: key material, transactions, hints and links. --json for the data
puzzles hints <id>	The collection's hints, the puzzle's own, then its hint files. --json for both
puzzles stages <id>	The stages of a multi-stage puzzle with their pages, files and published answers. --json for the list
puzzles assets <id>	The files a puzzle ships with SHA-256 and size, and the archived pages it cites. --read <path> writes one, --check/--live verify
puzzles list [collection]	One puzzle per line. --address, --chain, --status, --technique and --with-pubkey narrow it, --limit and --offset page it
puzzles verify [id]	A published key and its recipe against the address. --all or the list filters for a set, exit 1 on a mismatch
puzzles balance [id]	The live balance. The list filters check a whole set, one row each. --api-key, or one variable per chain
puzzles watch [id]	What the chain knows and the record doesn't. --since checks the source pages too, exit 1 on any finding
puzzles eligibility <query>	The checklist before working on a prize, by id or address. Exit 1 while any field is missing
puzzles export	The whole dataset with its data_version
puzzles mcp	The MCP server over stdio

--json prints the data behind the text, bigint as strings and absent fields left out. The commands come from the same tool definitions the agents get, so a typo in a flag fails here just like it fails for a model. The flags and exit codes are in the CLI guide.

🧠 Library

import { get, stats, verify } from "@agntn/puzzles";
import { bits } from "@agntn/puzzles/collections/bits";