18 AI tools for quantum developers

A circuit that passes a spec you wrote — target histogram, state or unitary, plus gate-set and depth limits — checked on the simulator, with every failed attempt and its diagnostics on the record.

Reproduces Verifier-in-the-loop code generation: the model proposes, an exact checker rejects, and the rejection is fed back — the pattern behind self-correcting LLM coding agents, applied to OpenQASM.

A MaxCut or state-preparation heuristic that beat its own ancestors — QAOA schedules and local-search programs proposed by several models, scored on the simulator, with the lineage of every survivor.

Reproduces Quantinuum's Hive (2025): a swarm of frontier models evolving Guppy and CUDA-Q programs against an execution harness, here reduced to a closed DSL and a visible budget.

A logical-vs-physical error-rate chart across the surface-code threshold, with a lookup table, an MLP and a transformer you trained in-browser racing exact MWPM at d = 3, 5 and 7.

Reproduces AlphaQubit (Bausch et al., Nature 2024): a transformer decoder that beat matching on surface-code memory experiments.

A drifting qubit whose X gate an RL agent holds above 99.5% from measurement counts alone, at fewer shots per step than recalibrating on a schedule — with the hidden truth drawn for you.

Reproduces Google's July 2026 result: an RL agent holding gate fidelity on Willow by adjusting 1,000+ control parameters from measurement statistics, replacing scheduled recalibration.

Your circuit routed onto a real coupling map twice — by SABRE and by a swap policy trained in your browser — with two-qubit count, depth and added SWAPs side by side.

Reproduces IBM's AI transpiler routing passes: 24% and 36% improvements over the default heuristic on 100+ qubit circuits, reproduced here with an eight-weight REINFORCE policy on 8–12 qubits.

A record your agent wrote over MCP while you slept: hypothesis, every circuit with its target and counts, conclusion.

Reproduces The hypothesis → run → record → conclude loop of agent-driven science (e.g. the AI Scientist, 2024), with targets and jobs modelled on Coda MCP, at the scale of a 16-qubit simulator.

The same circuit on your CPU and your GPU: wall time, the largest statevector each reaches under a stated budget, and the float32-vs-float64 gap.

Reproduces WebGPU statevector simulation — 2^N × 8 bytes of float32 amplitudes in a compute shader

Teleportation, superdense coding, Grover, the GHZ paradox, BB84 and E91 — each paper’s own protocol, executed live and checked against what the paper reported.

Reproduces Six landmark results (1984–1996) run on the site’s own presets and protocol simulators — not a general paper-to-circuit pipeline, a curated and citation-checked set.

A Clifford circuit on up to 4,096 qubits, run exactly in your browser: the stabilizer generators, the entanglement entropy across every cut, and measurement samples.

Reproduces The Gottesman–Knill theorem and the Aaronson–Gottesman tableau algorithm (2004) — the method behind Stim and every large-scale error-correction simulation.

A circuit run under a depolarising, T1/T2 and readout noise model, with the ideal, noisy, readout-mitigated and zero-noise-extrapolated expectation values side by side — error bars included.

Reproduces Zero-noise extrapolation by gate folding (Temme, Bravyi & Gambetta 2017; Li & Benjamin 2017) and readout inversion via a calibrated confusion matrix, as run on IBM and Google hardware.

An honest table: a variational quantum classifier against a tuned logistic regression and a small MLP on moons, circles and one quantum-generated set — held-out accuracy and training cost in circuit evaluations versus flops.

Reproduces The 2026 state of quantum machine learning: no accepted advantage on classical data, dequantised claims, data-loading and barren-plateau limits — and a demonstrated edge only when the data itself came from a quantum process.

A pass/fail verdict and process fidelity for whether two OpenQASM circuits are the same unitary up to global phase — paste both, get |Tr(V†U)|²/d² and a verdict.

Reproduces Unitary equivalence checking as used in compiler and transpiler validation — did an optimisation pass change what the circuit computes.

A correctness suite (Bell, GHZ-3, GHZ-5, 4-qubit QFT, Grover-2 vs. their closed-form targets) plus a performance sweep from 4 to 16 qubits, timed live in your own browser.

Reproduces The standard way a statevector simulator earns trust before it earns a speed claim: known-answer correctness tests, then a measured, reproducible throughput curve.

Every QPU qubit16 reaches — IBM, IonQ, Rigetti, Quantinuum — ranked for your circuit by projected success and cost per surviving shot, with each row’s calibration provenance and the model’s blind spots listed.

Reproduces Cross-vendor device selection: transpile to each native gate set, project whole-circuit survival from calibration medians, price it — the comparison no single-vendor console can make.

Physical qubits, runtime, code distance and magic-state factory count for an algorithm under surface-code error correction — plus how the answer explodes as the physical error rate nears threshold.

Reproduces Surface-code resource estimation (Fowler et al. 2012; Litinski 2019; Bravyi–Kitaev distillation) — the order-of-magnitude analysis behind results like Gidney’s 2025 RSA-2048 estimate.

Half-chain entanglement entropy against measurement rate for chains of 8 to 128 qubits, with error bars and CSV export — the volume-law to area-law transition, computed live.

Reproduces Measurement-induced phase transitions in hybrid random circuits (Li, Chen & Fisher 2018; Skinner, Ruhman & Nahum 2019), in the stabilizer-circuit form the original numerics used.

The [[n, k, d]] parameters, logical operators, degeneracy and single-error syndrome table of any stabilizer code you write down — with nine named codes to start from.

Reproduces The stabilizer formalism (Gottesman 1997): code parameters by symplectic linear algebra over GF(2), with each preset linked to its Error Correction Zoo entry.

Gradient variance against qubit count for random parameterised circuits — deep versus shallow, global cost versus local — on one log-scale chart with error bars.

Reproduces Barren plateaus in variational training landscapes (McClean et al. 2018) and their dependence on the cost function (Cerezo et al. 2021).