compare_devices
Rank by CX error, queue depth, qubit count, or combined score — only ranks devices whose real status is "active"; anything else is reported separately in unavailabledevices, never accidentally top-ranked
How to use it
compare_devices is exposed by the Quantum Hardware MCP MCP server. Add the server to your MCP client (Claude Desktop, Cursor, Windsurf and others), and the compare_devices tool becomes available to the model automatically. See the full listing for setup details and every tool this server provides.
Install Quantum Hardware MCP
/absolute/path/to/.venv/bin/python /absolute/path/to/quantum-hardware-mcp/server.pyOther tools in Quantum Hardware MCP (39)
Builds the qubit Hamiltonian from atom positions and reports how far entanglement forging and Pauli grouping cut the problem down — plus the accuracy floor at each truncation rank
Score and rank qubits by calibration quality — actually searches for a connected subset of the requested size, not just the top-n by score (falls back to warn-only if no connected subset of that size exists at all)
Emits the actual OpenQASM circuits for a forged ground-state calculation — each acting on half the qubits the molecule would otherwise need — plus a simulator self-check confirming they reconstruct the right energy
Cancel a queued or running job
Detects a silent hardware swap or qubit relabeling — the physical chip behind a device name changed, or its qubit indices got reassigned, neither of which any public API states directly. Built on a real per-qubit/per-pair calibration archive (qubitsnapshots/pairsnapshots), backfilled from IBM's own
Input: qubit interaction pairs → detects degree>3 nodes → predicts SWAP flood and gate inflation before it happens. Learned from Phase 4: degree-4 node caused 263→1,037 gate explosion.
Full dry-run: gate counts, qubit mapping, per-pair CX errors, estimated fidelity
Fetches the finished jobs and reconstructs the molecular energy, compared against the exact classical answer
Pre-submission check: missing measurements, decoherence violations, qubit mismatches
Calibration snapshots over the last N days
Exact calibration state on any past date — for paper reproducibility
Takes a value + sieve positions, builds one LNAA rail per k-column, searches all simultaneously in one hardware job
Input: Boolean conditions like {"1":1, "4":0} → derives Ising hi and Jij coefficients with full sign derivation and QAOA circuit recipe. The math behind Phase 5's 27.78× result.
Two-register LNAA — amplifies (n1, n2) pairs matching a Lucas mod-2 parity oracle (cross-register RZZ), without being told which rows to look for. Parity match is a weak, ~50%-hit-rate filter, not proof of equality — classical post-processing (comb()) checks every measured pair for true equality. Co
Bounds the error on any observable from one fidelity number, so you can tell before running whether a job can possibly reach chemical accuracy
Estimator primitive: computes ⟨ψ\
Predicts transpiled gate count from logical gates + max qubit degree. Knows the empirical ~600-gate noise floor on ibmmarrakesh.
QPU minutes + queue wait estimate before you submit
Curve intersection search — integer root-finding across column pairs to jump directly to candidate rows
Calibration drift alerts — spikes >20% in CX error, readout error, T1, or T2. The T1/T2 check reads live from the shared Turso database (2026-08-30) when configured, same one quantum-verifier uses, so this reflects real current state regardless of whose laptop last synced — falls back to the local d
Input: job ID + marked bitstrings → amplification factor, per-state shot breakdown, verdict (EXCELLENT/GOOD/WEAK/FAILED).
Per-qubit T1/T2, readout error, gate error, queue depth
Bit-string measurement counts from a completed job
QUEUED / RUNNING / DONE / ERROR
All accessible IBM backends with live operational status
Recent jobs with status, backend, and timestamps
Given a molecule and a budget, works out the most accurate result you can actually buy: Schmidt rank, circuit count, cost, expected error
Current queue snapshot across all backends
Which mitigation techniques are worth applying for a given circuit and device noise — including the ones measured not to help, so you skip them
KL-divergence reproducibility score (0 = identical, 1 = maximally different)
Credit-aware routing — cheapest backend that meets your error threshold
Builds the circuits and submits them to a named device via submitjob, returning ordered job IDs
Full Grover's search — builds oracle + diffusion operator, picks least-busy backend, submits
Variational Quantum Eigensolver — H2 ground state to chemical accuracy
Classical Lucas theorem sieve — filters 98%+ of Pascal's Triangle search space before touching the QPU
Run the same circuit N times, record variance across runs
Transpile and submit OpenQASM 2.0 or 3.0 — returns jobid
Exact measurement distribution for any Clifford-only circuit (H, S, CX, CZ, ...) via the stabilizer tableau — not simulated, not estimated, exact, and scales to hundreds of qubits (Gottesman-Knill theorem). Confirmed: a 150-qubit Clifford circuit verifies in under a second, where state-vector simula
Verifies real hardware measurement counts against a Clifford circuit's exact stabilizer prediction — a real fidelity lower bound at any qubit count, no simulation required