UQH · Universal Quantum Hub
Beta — open to use, surface still moving

Universal Quantum Hub

Managed quantum infrastructure for research institutions, national programs, and enterprise R&D. Connect every framework and vendor, run the full pipeline under governance and cost control, and read the hardware landscape in data — no vendor lock-in, no rewrites per provider.

Works with your SDK
QiskitPennyLaneQrispCUDA-QCirqQadencePercevalPulserBraket-AHS
Multi-vendor by design
IBM QuantumIQMIonQRigettiQuEraPasqalQuandelaQuantinuum
8
Hardware vendors
9
SDK frameworks
7
Independent pipeline steps
4
Application modules
The platform

One middleware, from circuit to confident result

Bring your circuit once. UQH detects the framework, optimizes and noise-checks it, applies error mitigation and error correction, prices it across vendors, and runs it on the backend you choose — then turns every run into operational intelligence.

Connect everything

Nine SDK frameworks and eight hardware vendors across superconducting, trapped-ion, neutral-atom, and photonic — behind one interface.

Run with confidence

Noise simulation, readout + ZNE mitigation, Pauli-check postselection, and QEC analysis make results trustworthy — with cost known before you submit.

Operate at scale

Accumulated jobs become benchmarks and device intelligence; per-domain policy, billing, and an AI-agent interface make it run as a service.

Applications

Ready-to-run quantum application modules

Domain modules you can run today — each comparable across CPU, GPU, and real QPU backends.

QML anomaly detection

Quantum-kernel outlier detection benchmarked against a classical baseline — upload your own CSV; advantage is dataset-dependent.

Detection quality · AUC
QuantumClassical0.920.88

Quantum finance

QAOA portfolio optimization versus exact and greedy baselines.

Approx. ratio · CPU / GPU / QPU
CPUGPUQPU1.001.000.97

Quantum chemistry

VQE ground-state energy with noise and mitigation comparison.

VQE dissociation curve · H₂
equilibriumbond length →

Graph optimization

QAOA for max-cut, independent set, and vertex cover.

Max-Cut · 6-cycle
cut = 6 (optimal)
Before you spend a shot

QPU time is scarce. Spend it on the run that will work.

Most wasted quantum spend is decided before submission — the wrong device, a circuit too deep for it, or a run you already made last month. UQH answers all three up front.

Analytic fidelity estimate

An error-budget calculation from the device's measured 1q, 2q and readout error rates against your circuit's gate counts. No noise simulation, so no qubit ceiling and no wait — it answers for a 100-qubit circuit as fast as a 2-qubit one.

Rank the devices, vendor-neutral

The same circuit scored across every connected processor — expected fidelity, 2q error, coherence headroom, queue depth and cost, side by side. Pick on evidence, not on which vendor you happen to have an account with.

You may have run this already

Submissions are matched by circuit content, not filename, so earlier runs of the same circuit surface before you pay for another one — and an edited circuit correctly stops matching. Re-running is a choice, not an accident.

Prove the circuit on GPU first

Run it on dedicated GPU simulation hosts before it ever reaches hardware. Logic bugs, wrong observables and bad ansätze are cheaper to find on a GPU than in a QPU queue.

Workbench

One workspace, from upload to result

Drop in a circuit and UQH detects its framework, previews the metadata, and runs the whole pipeline — analyze, optimize, noise-check, mitigate, submit — without leaving the page.

  • Framework auto-detection on upload — Qiskit, PennyLane, Qrisp, and more.
  • Favorites, recents, and inline circuit-metadata preview.
  • Run any single step or the full pipeline from the UI.
  • Emulator dry-run to validate before you spend on real hardware.
Pipeline · bell.py
  • analyzedone
  • optimizedone
  • noise simdone
  • mitigatedone
  • submitready
Capabilities

Everything the quantum workflow needs

A broad, composable feature set — use any piece on its own, or the whole pipeline end to end.

Frameworks & vendors

Write in your SDK of choice and submit to any supported processor — no per-vendor rewrites.

QiskitPennyLaneQrispCUDA-QCirqQadencePercevalPulserBraket-AHS

Independent pipeline

Analyze, optimize, noise-simulate, QEC, GPU-execute, QPU-submit — each step runs standalone, with an emulator dry-run option.

analyzeoptimizenoiseQECrun

Circuit optimization

Transpile and route with multiple engines, mapped to the target device's physical layout — and compare every option side by side on resulting complexity, optimization time, and analytical fidelity. Every candidate is verified mathematically equivalent to your original circuit — large circuits offloaded to GPU — and comparisons run in isolation, so timings stay honest and the platform stays responsive.

QiskitTKETQuiZXBQSKitSABREverified

Error mitigation & resilience

Readout-error mitigation and zero-noise extrapolation as one resilience level, plus spacetime Pauli-check postselection for Clifford circuits — standard techniques, applied consistently.

Noise simulation & QEC

Calibration-based fidelity and TVD with a cross-engine consistency check (Qiskit vs PennyLane), plus error-correction analysis and code application with threshold and logical-error estimates.

fidelity/TVDsurfacerepetition

Cost & resource estimate

See qubits, depth, gates, and what each vendor bills — in that vendor’s own unit — before submitting. No surprise spend.

GPU acceleration

Offload heavy simulation to dedicated GPU hosts and compare CPU, GPU, and QPU on the same problem.

CUDAlightning.gpu

Knowledge base (RAG)

Ask in your own words — that is the whole interface. Hybrid dense + sparse retrieval, re-ranking, and cited answers over the platform's own accumulated runs: measured results, refusals and workarounds, not a document corpus. Findings distilled from them are anonymised and searchable by everyone; individual records stay with their owner. What the corpus contains is curated by your administrators, not chosen per query.

API, CLI & MCP

Full parity across UI, REST API, and CLI — plus an MCP server so AI agents orchestrate runs directly.

RESTCLIMCP

Self-maintaining (AutoOps)

A scheduled autonomous layer tracks every device's fidelity and error rates over time and flags the ones drifting away from spec, watches SDK releases against your pinned versions, and picks up new hardware as vendors bring it online. Reversible actions auto-apply and log; anything irreversible waits for admin approval.

Hybrid-ready execution

Algorithm files already run on remote compute hosts, and the backend a program names is swapped at run time — a script written for a simulator reaches a GPU host or real hardware without being edited. When tightly coupled GPU-QPU machines arrive, they join as one more target on that same path rather than as a new architecture.

Reproducible results

Export one self-describing file per run: the circuit, the observable, the calibration exactly as it stood at submit time, the result, the noiseless reference and SDK versions. Anything unavailable is named rather than quietly omitted, so the number stays checkable years later — attach it to a paper or hand it to a reviewer.

Coming soon

Runtime sessions

Reserve a device and batch many jobs in one session to minimize re-queuing between iterations — in the spirit of vendor runtime sessions. On the roadmap.

Intelligence

Read the ecosystem in data, not hype

Every job you run feeds a living view of the hardware landscape — so researchers and decision-makers act on measured reality.

QPU Fidelity Leaders · fidelity
ibm_fez0.90
iqm_emerald0.74
iqm_garnet0.61
rigetti0.55
Device health · drift over time
fidelitytime →
Cost vs fidelity · per device
fidcost →
Playbooks · best device per circuit
  • ≤10 qubitsibm_fez
  • 11–20 qubitsiqm_emerald
  • 21–30 qubitsiqm_sirius
SDK Leaders · composite
Qiskit94
PennyLane83
CUDA-Q64
Cirq52
Optimizer Leaders · depth reduction
TKET13%
Qiskit L30%
Reuse or re-submit · calibration drift
  • Cal stable (<15% drift)reuse result
  • Cal drifted (≥15%)re-submit
Every run stores the device calibration — compare it to now and skip the spend when nothing moved.
Workflow & developers

One pipeline — the same power in your terminal

Every step is independent and scriptable. Run it in the UI, call the REST API, or drive it from the CLI and your CI.

analyze
optimize
noise sim
QEC
GPU exec
QPU submit
uqi — zsh
$ uqi estimate --file bell.py
  ibm_fez   $0.00     iqm_garnet  $0.30
  rigetti   $0.74     ionq_forte1 $82.22   # cheapest first

$ uqi catalog optimize --graph cycle6 --compare --resilience full
  CPU optimal   ·   GPU optimal
  QPU(iqm_emerald)  fid 0.74 · valid 100%
  readout-mitigated  approx 100%   # post-processed, no re-run
Agent mode

Let an AI agent run your quantum workflow

UQH ships an MCP server, so Claude and other AI agents use it as a tool — analyzing, optimizing, mitigating, pricing, and submitting across vendors from a single natural-language request.

  • Natural-language workflow — describe the goal, the agent runs the steps.
  • Every step is a tool — analyze, optimize, mitigate, estimate, submit, compare.
  • Multi-vendor orchestration — the agent selects and runs across backends.
  • Standard MCP — works with Claude and other MCP clients.
agent · MCP
You: optimize bell.py, pick the cheapest backend, then mitigate
agent
   uqi_analyze            2 qubits · depth 2
   uqi_resource_estimate  ibm_fez $0.00 (cheapest)
   uqi_optimize           SABRE-routed
   uqi_qpu_submit         submitted 
   uqi_mitigate_readout   approx 100%
Hardware

Every modality, one platform

Reach superconducting, trapped-ion, neutral-atom, and photonic processors — and compare and run across them from a single place.

VendorModalityQubitsAccess
IBM QuantumSuperconducting133+Open plan
IQMSuperconducting20–150Credits
IonQTrapped-ion36AWS Braket · IonQ Cloud
RigettiSuperconducting108AWS Braket
QuEraNeutral-atom256AWS Braket
PasqalNeutral-atom100PCS · Azure
QuandelaPhotonicFree tier
QuantinuumTrapped-ion20–56Azure · HQC

Which SDK reaches which hardware

Write in the framework you already use. Gate-based circuits reach every gate-based vendor; photonic and analog work stays with the hardware that can express it.

SDKIBMIQMIonQRigettiQuEraPasqalQuandelaGPU sim
Qiskit···qiskit-aer GPU
PennyLane···lightning.gpu
Cirq···qsimcirq
Qrisp···qiskit-aer GPU
Qadence···torch CUDA
CUDA-Q···nvidia target
Perceval·······
Pulser·······
Braket-AHS·······
Who it's for

Built for institutions and government — open to every researcher

Institutions & government

The primary fit — national programs, research institutes, and enterprise R&D. One governed, white-label control plane over every vendor: spend safeguards, code and data isolation, curated knowledge scope, security audit of every refused upload, and data sovereignty — with Leaders boards and real cost-vs-fidelity to ground hardware and budget decisions.

Researchers

Individual researchers aren't shut out — everything up to the point of spending money is self-serve and needs no credential, and once you register your own vendor key the same workspace runs it: compare CPU, GPU and real QPUs, watch noise appear, and recover signal with mitigation.

Developers

Work from the environment you know. Circuit-complexity analysis, optimization, cost estimates, and runs over API and CLI — script them or wire them into CI.

Security & governance

Built to run as a managed service

Multi-tenant controls, spend safeguards, and isolation are enforced server-side across UI, API, CLI, and MCP.

Per-domain white-labelControl visible tabs, pipeline steps, devices, and accounts per domain and plan.
Spend safeguardsPre-submission estimates in each vendor’s own billing unit, a cost threshold, and a per-job usage record. Work runs on your own vendor account — we never hold a balance for you.
Code & data isolationStatic analysis and execution limits on uploads; per-user file and job isolation.
Auth & privacySession and API-key auth, output escaping, and PII scrubbing before external LLM calls.
Abuse defensePer-IP flood auto-blocking and IP/account blacklist, enforced across UI, API, CLI, and MCP at a single server chokepoint.
Institutional governanceMembers register their own vendor credentials, so one person’s allowance is never spent by someone else. Institutions manage membership and access; every admin action is audit-logged.
Institutional reportingOn request, export a full account of an institution's activity — every member's quantum, AI and GPU work, with what it cost in each vendor's own unit — over any period or all time, scoped strictly to that institution and carrying its own audit trail.
Quantum, AI and GPU in one recordQuantum time, AI calls and GPU runs land in the same institutional record, so one export shows where the work went — no separate accounting for the classical half. Each still bills where it belongs: quantum and AI on the member's own vendor account, GPU on hardware the institution attached and opened to its members.
Credential vaultVendor keys are held encrypted under the account that registered them, and nowhere else — there is no institution key and no platform key to lend, because a vendor allowance is issued to a person. A key can be limited to named devices, and a compromised one is suspended with a reason rather than deleted, so its owner learns why instead of watching it vanish.
Access models

Three ways to run, one platform

Everything that costs nothing to run is open to everyone, with no credential at all. What costs money — quantum processor time, AI calls — runs on credentials you bring yourself. There is no shared platform allowance to draw on: a vendor allowance is issued to a person, so lending ours would breach that vendor's terms for the holder.

No credentialsEverything that costs nothing is yours immediately: analysis, optimization, noise simulation, QEC, device ranking, reproducibility bundles, and the API, CLI and MCP interfaces. Submitting to a real QPU and using AI features are the two things that need a key of your own.
Bring your ownRegister your vendor account key and runs execute on it. The vendor enforces its own limit on its own key and bills you directly — we hold no balance for you and never fall back to a platform key.
InstitutionMembers register their own credentials, as everyone does. What the institution governs is who belongs, what they may reach, and any GPU hardware it attaches and opens to them — with a cap on what one submission may cost, reporting over any period, and an audit trail.
CapabilityNo credentialsBring your ownInstitution
Analysis, optimization, noise simulation, QEC
Vendor-neutral device ranking
Reproducibility bundle
API, CLI and MCP access
Credential registration·
Restrict a registered key to named devices·
Submit to a real QPU·
Per-submission cost cap and audit trail··
Institutional usage report··
Bring your own AI key·
GPU simulation on shared hosts
Unmetered use of a GPU host you provide·
Why UQH

The combination is the difference

Individually, these capabilities exist elsewhere. UQH brings them into a single flow.

QEC-aware pipelineError correction sits in the same flow as optimization and execution.
Agent-native executionMulti-vendor runs orchestrated directly by AI agents over MCP.
Result-quality evaluationFidelity and benchmarking grounded in your own job history.
Per-domain policyWhite-label control over features, devices, and access by domain and plan.