Entropy

Verifiable randomness for AI agents. Four tools, no auth, free forever.

Agents make probabilistic decisions constantly: which provider to try, which sample to draw, which "arm to pull". Almost none of it is auditable. Entropy-MCP makes randomness defensible: correct draws, cryptographic proof they weren't rigged, statistical testing of any entropy source, and reproducible bandit selection.

Tools:

1. random: correct draws, optionally reproducible

Six operations: bytes, int, shuffle, choose, sample, constrained.

The implementations are the point. Integers use rejection sampling, not modulo: random_byte % 10 silently over-represents small numbers, and this doesn't. Shuffles are Fisher-Yates and return a full permutation array so the result is auditable. Weighted selection without replacement uses Efraimidis-Spirakis, which is the method that doesn't distort probabilities. sample draws from six named distributions (uniform, normal, lognormal, exponential, triangular, beta) with a summary block attached.

2. commitment: prove the draw was fair

Full commit-reveal lifecycle: commit, reveal, draw, verify, list.

Commit before the draw: the server stores only the SHA-256 of the seed, binds it to your draw spec, timestamps it against an NTP-verified clock, optionally anchors it to a drand beacon round, and attests the record. The seed goes to you and is never stored server-side. Reveal hands it back, verifies the hash, re-executes deterministically, and attests the outcome.

verify is stateless and portable. Hand it a commitment and a reveal from any instance and it checks seven things: seed-to-hash, both attestations intact, commit preceded reveal, outcome reproduces, spec unaltered, beacon anchor genuine. No local state required, which means any third party can audit any draw.

3. test: is this source actually random?

Eight tests in the spirit of NIST SP 800-22 (no certification claimed): frequency, chi-squared, runs, longest-run, serial, Shannon entropy, Kolmogorov-Smirnov, and gap. Accepts numbers, categories, or hex/base64 bytes. Each test returns a statistic, p-value, pass/fail at your alpha, and a plain-English reading.

Built-in honesty: at alpha=0.05 a battery fails roughly one test in twenty on perfectly sound entropy by chance alone. The output says so.

4. explore: which arm do I pull?

Stateless multi-armed bandit selection: epsilon_greedy, thompson, ucb1. You own the state and pass arm history in; you get back an arm and the reason. Ties break randomly rather than by list order, because index-order tie-breaking is itself a bias. Pass a seed and the selection replays exactly — which matters when the selection is part of a decision someone will later question.

Why it exists

Randomness in production AI systems is usually a black box: a call to a library, an outcome, no record. That's fine until someone asks whether the assignment was fair, whether the synthetic data matched its spec, whether the RNG you inherited is sound, or why the agent picked that provider.

DISCLAIMER: Entropy queries public Python libraries. Implementers are responsible for what they put in payloads, Entropy doesn't log payload content. Usage in production systems is at the implementer's risk.