Agentic layer for Quantum Software

Simulate and compare circuits, catch regressions, resume optimization runs, and share recorded results. Review exactly what will be submitted before running on hardware.

Get started Watch a session

Circuit files today. Python scripts and notebooks next. Qiskit native.

One session

From circuit check to recorded result.

The plan is printed before anything runs. The approval is for that plan's digest. The result points at an evidence file you can hand to someone else.

skaden — bell-state/

$ skadenskaden v0.9.0 · mode=ask · sandbox=seatbelt · adapter=qiskit-local > check bell.qasm against the saved baseline   plan     circuit_check(path=bell.qasm, baseline=.skaden/baselines/bell.json)  scope    read-only · simulator · no network  digest   sha256:9f2c41ab…7e03   approve this exact call? [y/N] y   ✓parsed      2 qubits · 3 gates · depth 2  ✓simulated   1024 shots · seed 7 · reproducible  ✓equivalent  operator method · baseline matched  ✓evidence    .skaden/regression/run_2f8c8890.json   PASS · no semantic drift from baseline $ skaden experiment reproduce run_2f8c8890  ✓reproduced  digest match · 0 divergences

plan
Shown before it runs
The exact call, its scope, and its digest.
approve
One digest, one approval
Edit the plan and this consent is void.
evidence
A file, not a claim
Inputs, seed, environment and result, chained.

Capabilities

Check your circuits
Simulate, draw, lint, compare, and check equivalence within supported limits. Block changes that exceed your regression thresholds.
Resume your experiments
Run VQE and QAOA with saved progress. Record seeds, shots, and software versions.
Prepare for hardware
Build calibration-based noise models, select qubits, and review submissions and estimated costs. IBM hardware access is in preview.
Share your results
Export a portable experiment bundle with a readable report and offline integrity checks.

Run and compare experiments

Evidence

A result others can inspect.

Inspect the exported report and verify bundle integrity without quantum software. Recomputing an experiment requires its supported software environment.

Experiment bundlePortable / inspectable

SourceInputs and settings: Inspect the circuit, experiment settings, seeds, and shots recorded for the run.
EnvironmentSoftware environment: See the recorded software versions needed to recompute the experiment.
RecordResults and integrity: Read the results and check that the files in the exported bundle are consistent.

Regression gate

Circuits get a CI gate. Inconclusive is a real verdict.

Save a reviewed baseline. Later versions are checked for structural drift, bounded equivalence and sampled-distribution distance. The model cannot change the verdict, and a check that cannot decide says so instead of passing.

circuit regressionInteractive example

Recorded verdict

pass

Every declared check held for the recorded source, engine, shots, seed and thresholds.

CI processexit 0

Declared checksAgainst baseline

Circuit depth
Passed
Baseline2Current2
Maximum increase: 0Unchanged from baseline.
Two-qubit gates
Passed
Baseline1Current1
Maximum increase: 0Unchanged from baseline.
Equivalence
Passed
Equivalent
Method: operatorOperator comparison completed.
Distribution distance
Passed
Total variation distance0.000
Maximum TVD: 0.1Total variation distance from the baseline distribution.

Inspect report.jsonRaw example

{  "schema": "skaden.circuit-regression/v1alpha1",  "verdict": "pass",  "depth":        { "baseline": 2, "current": 2, "max_increase": 0 },  "two_qubit":    { "baseline": 1, "current": 1, "max_increase": 0 },  "equivalence":  { "method": "operator", "result": "equivalent" },  "distribution": { "total_variation": 0.000, "max": 0.1 },  "digest": "sha256:2f8c8890…d5e9"}

Recorded verdict

fail

Depth and two-qubit gate counts exceed the baseline. CI exits with code 2.

CI processexit 2

Declared checksAgainst baseline

Circuit depth
Exceeded
Baseline2Current4
Maximum increase: 0+2 from baseline.
Two-qubit gates
Exceeded
Baseline1Current3
Maximum increase: 0+2 from baseline.
Equivalence
Passed
Equivalent
Method: operatorOperator comparison completed.
Distribution distance
Passed
Total variation distance0.004
Maximum TVD: 0.1Total variation distance from the baseline distribution.

Inspect report.jsonRaw example

{  "schema": "skaden.circuit-regression/v1alpha1",  "verdict": "fail",  "depth":        { "baseline": 2, "current": 4, "max_increase": 0 },  "two_qubit":    { "baseline": 1, "current": 3, "max_increase": 0 },  "equivalence":  { "method": "operator", "result": "equivalent" },  "distribution": { "total_variation": 0.004, "max": 0.1 },  "failed": ["depth", "two_qubit"]}

Recorded verdict

inconclusive

The equivalence engine hit its bound. Nothing is guessed. CI still goes red.

CI processexit 2

Declared checksAgainst baseline

Circuit depth
Passed
Baseline41Current41
Maximum increase: 0Unchanged from baseline.
Two-qubit gates
Not reported
Not reported
This example report does not include a two-qubit check.
Equivalence
Bound exceeded
Inconclusive
Method: operator26 qubits exceeds the bounded operator engine
Distribution distance
Passed
Total variation distance0.012
Maximum TVD: 0.1Total variation distance from the baseline distribution.

Inspect report.jsonRaw example

{  "schema": "skaden.circuit-regression/v1alpha1",  "verdict": "inconclusive",  "depth":        { "baseline": 41, "current": 41, "max_increase": 0 },  "equivalence":  {    "method": "operator",    "result": "inconclusive",    "reason": "26 qubits exceeds the bounded operator engine"  },  "distribution": { "total_variation": 0.012, "max": 0.1 }}

View CLI commandCLI command

skaden check bell.qasm \
  --baseline .skaden/baselines/bell.json \
  --equivalence-method operator \
  --max-depth-increase 0 \
  --max-total-variation 0.1

Claim boundary

A pass means the declared checks passed. It does not prove algorithmic correctness, quantum advantage or hardware performance.

Roadmap

What comes next

Each release is complete when researchers can perform its workflows independently on their own machines.

Available now: Circuit workflows
Run and check circuit files, resume supported optimization runs, prepare hardware submissions, and export experiment evidence.
Release 1: Rerunnable research
Record existing Python scripts and notebooks, rerun supported experiments, and investigate differences across inputs and software versions.
Release 2: Approved experiment automation
Describe an experiment, approve its plan, and run recorded parameter sweeps and device-aware simulations.
Release 3: Hardware and long studies
Add Amazon Braket execution, resume studies across days, and compare raw and error-mitigated results with uncertainty estimates.
Release 4: Group workflows
Reuse research history, add native Cirq and PennyLane support, and share recorded experiments across your group.

Execution controls

You control what runs

Review plans before consequential execution. Hardware submission uses a separate, human-controlled path. Provider credentials stay outside the model's access. Decisions and results are recorded.

IBM hardware preview

Two human confirmations are required before hardware submission.

Read the threat model Execution details

Backends

Hardware agnostic. Labelled honestly.

One versioned adapter SDK with registration, health checks and conformance tests. What is not built yet is marked as not built yet.

nowshipped
previewopt-in preview
sooncatalog or roadmap

Backend and adapter status
Backend	Status	What that means today
Qiskit + Aer	Native	Inspect, compile, draw, lint, diff, estimate, translate, simulate noise, and check seeded reproducibility.
OpenQASM 2/3	Native	Bounded, digest-bound intake with guarded includes and parse validation before anything is approved.
External MCP	Remote	Streamable HTTP servers attach through the same gateway. Local stdio stays off until it can be isolated.
IBM Quantum	Protected preview	Live targets, immutable plans and human-only submission through the broker. No credentialed device canary yet.
AWS Braket	Catalog only	Dated target metadata. Planning and execution adapters are roadmap.
Azure QDK	Catalog only	No Q# compiler or workspace adapter yet.
Cirq + PennyLane	Roadmap	Project detection and OpenQASM export intake today. Native adapters next.

Qiskit + Aer: Native; Inspect, compile, draw, lint, diff, estimate, translate, simulate noise, and check seeded reproducibility.
OpenQASM 2/3: Native; Bounded, digest-bound intake with guarded includes and parse validation before anything is approved.
External MCP: Remote; Streamable HTTP servers attach through the same gateway. Local stdio stays off until it can be isolated.
IBM Quantum: Protected preview; Live targets, immutable plans and human-only submission through the broker. No credentialed device canary yet.
AWS Braket: Catalog only; Dated target metadata. Planning and execution adapters are roadmap.
Azure QDK: Catalog only; No Q# compiler or workspace adapter yet.
Cirq + PennyLane: Roadmap; Project detection and OpenQASM export intake today. Native adapters next.

Install

Start with your next quantum experiment.

One pinned commit, one command. Everything simulates locally until you say otherwise.

Get started View on GitHub

# pin a reviewed commit, then:
uv tool install --python 3.12 \
  --from "skaden[qiskit] @ git+ssh://..." \
  skaden

Needs uv, Git and access to the repository. There is no PyPI release yet. The install pins one reviewed commit on purpose. Full install guide →

REVISION=REPLACE_WITH_REVIEWED_40_CHAR_COMMIT_SHA
uv tool install --python 3.12 \
  --from "skaden[qiskit] @ git+ssh://git@github.com/skadendev/skaden.git@${REVISION}" \
  skaden

Agentic layer for Quantum Software

Shown before it runs

One digest, one approval

A file, not a claim

Check your circuits

Resume your experiments

Prepare for hardware

Share your results

pass

Circuit depth

Two-qubit gates

Equivalence

Distribution distance

fail

Circuit depth

Two-qubit gates

Equivalence

Distribution distance

inconclusive

Circuit depth

Two-qubit gates

Equivalence

Distribution distance

Available now: Circuit workflows

Release 1: Rerunnable research

Release 2: Approved experiment automation

Release 3: Hardware and long studies

Release 4: Group workflows

Start with your next quantum experiment.