QECTOR is a local, fail-closed quantum error correction (QEC) integration
for Claude Code and Claude Desktop, built on qector-decoder-v3==1.0.0
and the QECTOR Decoder v3 Reference Manual (DOI: 10.5281/zenodo.21941046).
Every correction returned by this plugin is verified against
H c = s (mod 2) before it leaves the local server, and the default
operation makes no network request.
4 MCP servers · 8 stable / 29 research / 3 admin tools · 11 commands · 5 agents · 28 skills · zero-egress default
1. Install the runtime
git clone https://github.com/GuillaumeLessard/qector-claude-plugin.git
cd qector-claude-plugin
python -m pip install -r requirements.txt
python scripts/qector_runtime_check.py2. Install as a Claude Code marketplace plugin
claude plugin marketplace add GuillaumeLessard/qector-claude-plugin
claude plugin install qector@qector-toolsThe default install exposes only qector-library (8 stable tools). Opt-in
research and admin surfaces are added explicitly — see the trust-zone
table below.
3. Install the local Claude Desktop extension (Windows / macOS / Linux)
The Claude Desktop manifest lives under .claude-desktop-extension/ and
points at mcp/mcp_server_desktop.py --profile safe. The local installer
records the exact Python interpreter used, so Windows PATH ambiguity
cannot break the install:
python scripts/configure_claude_desktop.py --check-only # dry run
python scripts/configure_claude_desktop.py --confirm # write configA native Windows command is also shipped:
.\scripts\install_windows_connector.cmd| Server | Purpose | Tools | Default |
|---|---|---|---|
qector-library |
Frozen stable decoding surface | 8 | ✅ |
qector-research |
Provisional research / evidence tools | 29 | opt-in |
qector-admin |
Privileged local operations | 3 | opt-in |
qector-desktop-mcp |
Claude Desktop safe profile | 8 | ✅ (Desktop) |
qector-library is the authoritative default surface. Every tool on it
verifies its output before returning: syndrome decoders check
H c = s (mod 2), threshold sweeps ship Wilson 95% intervals with
SHA-256-stamped JSON artifacts, and build_code_from_matrix validates
dimensions and code family before allocating a parity-check matrix.
The research and admin servers are never auto-enabled. They are added
only by an explicit --with-research or --with-admin flag combined
with a confirm=true argument on every call.
/qec-setup, /qec-facts, /qec-theorem, /qec-reproduce,
/qec-threshold-sweep, /qec-wilson, /qec-dem, /qec-code-inspect,
/qec-benchmark, /qec-sinter, /qec-validate-mcp.
qec-researcher, qec-developer, qec-validator, qec-sysadmin,
qec-hardware-engineer.
qector-core, qector-architecture, qector-math-foundations,
qector-codes-builder, qector-bp-osd, qector-two-stage-css,
qector-space-time, qector-decoders-deep-dive, qector-dem-pipeline,
qector-sinter, qector-ler-methodology, qector-pymatching-compat,
qector-batch-decoding, qector-orchestration, qector-reproducibility,
qector-testing-strategy, qector-deployment, qector-release-engineering,
qector-licensing, qector-services, qector-glossary, qector-educator,
qector-researcher, qector-developer, qector-sysadmin,
qector-hardware-engineer, qector-roadmap, qector-workbench.
The 40 tools split into three named trust zones, plus the Desktop safe profile which re-exports the 8 stable tools through a separate adapter. No tool calls the network by default.
| Zone | Count | Default? | Enables on | Risk posture |
|---|---|---|---|---|
| Stable (library) | 8 | ✅ | always | every result is verified |
| Research (bench) | 29 | opt-in | --with-research |
read / compute; Wilson CIs; SHA-256 artifacts |
| Admin | 3 | opt-in | --with-admin + QECTOR_ADMIN_ENABLED=1 |
confirm=true required per call |
The MCP processes enforce their own ceilings because there is no external
rate limiter. The defaults are intentionally conservative; every ceiling
is overridable per-tool with QECTOR_MCP_MAX_CALLS_<TOOL_NAME> and
disable-able with -1.
| Tool | Default | Why it is limited |
|---|---|---|
threshold_sweep |
8 | writes hashed artifacts; large LER grid |
decode_single |
64 | seeded decode loop; can be retried |
decode_syndrome |
256 | per-call decode compute |
build_code_from_matrix |
32 | arbitrary parity-check matrix |
hot_path_microbench |
4 | machine-scoped; no portable claim |
system_setup |
2 | can install packages, write dirs |
configure_claude_desktop |
2 | rewrites Claude Desktop config |
workbench_probe |
2 | launches a local executable |
Exhaustion returns RESOURCE_LIMIT and the process stays up. See
SECURITY.md for the input-size ceilings
(QECTOR_MCP_MAX_DISTANCE, QECTOR_MCP_MAX_TRIALS,
QECTOR_MCP_MAX_SWEEP_POINTS, QECTOR_MCP_MAX_MATRIX_CELLS).
- It does not transmit syndromes, parity matrices, circuits, or tool
arguments anywhere by default. The only outbound network operation is
the explicit opt-in PyPI freshness check
(
compat_report(check_pypi=true)/env_block(check_pypi=true)); it checks the published package version only and caches the response for the server process. SeePRIVACY.md. - It does not run untrusted code. The three tools classified as
"Execute" (
decode_single,threshold_sweep,build_code_from_matrix) perform scoped QEC math on caller-supplied binary inputs. They never spawn subprocesses, never load dynamic modules, and never read or write outsideQECTOR_ARTIFACT_DIR. - The four "Write" tools are limited to: writing the matrix file the
caller just built (
code_export_matrices), writing the local Claude Desktop config (configure_claude_desktop, requiresconfirm=true), and writing a hashed artifact under the configured artifact directory (artifacts_sha256/artifact_metadata_check). None touch arbitrary filesystem paths or remote endpoints.
QECTOR Claude Plugin v1.0.4
┌──────────────┐ ┌──────────────┐ ┌────────────────┐
│ 11 Commands │ │ 5 Agents │ │ 28 Skills │
└──────┬───────┘ └──────┬───────┘ └────────┬───────┘
│
▼
┌──────────────────────────────────────────────┐
│ qector-library MCP 8 stable tools │ default
│ qector-research MCP 29 provisional tools │ opt-in
│ qector-admin MCP 3 privileged tools │ opt-in
│ qector-desktop-mcp 8 safe-profile tools │ Desktop
└──────────────────────┬───────────────────────┘
│ local stdio, zero egress
▼
┌────────────────────────┐
│ qector-decoder-v3 │
│ 1.0.0 (Rust/PyO3) │
└────────────────────────┘
The four servers share mcp/qector_mcp_contract.py, which enforces the
result envelope (status, tool, server, version), the closed
list of stable error codes, and the fail-closed default for every
operation.
The canonical 1.0.4 release assets live in dist/, each with a
SHA-256 sidecar and a combined SHA256SUMS file:
| Artifact | Contents |
|---|---|
qector-claude-plugin-1.0.4.zip |
Claude Code plugin: skills, agents, commands, prompts, hooks, MCP servers, docs |
qector-claude-plugin-source-1.0.4.zip |
Public QECTOR source distribution (no test suite, no CI workflows, no internal decks) |
qector-claude-desktop-1.0.4.mcpb |
Claude Desktop safe-extension MCPB (8 stable tools, icon.png and README.md at bundle root) |
qector-claude-plugin-1.0.4.sbom.json |
SPDX-2.3 SBOM for the three packages above |
provenance.json |
per-artifact SHA-256 + git commit + runtime pin |
SHA256SUMS |
combined sidecar file |
qector-claude-desktop-1.0.4.mcpb is the artifact listed in
server.json and is the one Claude Desktop's MCP Registry downloads.
| Component | Requirement |
|---|---|
| Python | 3.9 – 3.13 (tested on 3.12) |
| OS | Windows, macOS, Linux |
| Backend | qector-decoder-v3==1.0.0 |
| MCP runtime | mcp==1.26.0 |
| Scientific stack | numpy>=1.26,<2.3, cryptography>=48.0.1,<50 |
| Network | none by default; one opt-in PyPI check |
| GPU | none — no portable speed claims are made |
| Document | Path |
|---|---|
| User Manual | docs/User_Manual.md |
| MCP API | MCP_API.md |
| Security | SECURITY.md |
| Provenance | PROVENANCE.md |
| Privacy | PRIVACY.md |
| Release Validation | RELEASE_VALIDATION.md |
| Tool Stability | TOOL_STABILITY.md |
| Claude Desktop Notes | CLAUDE_DESKTOP.md |
| Changelog | CHANGELOG.md |
| Reference manual (v1.0.0) | DOI 10.5281/zenodo.21941046 |
python -m unittest discover -s tests -v # mathematical + protocol tests
python scripts/validate_source.py # source-only structural check
python scripts/validate_plugin_bundle.py # built-bundle check
python scripts/release_validate.py # version + manifest cross-check
ruff check . # lint
python scripts/build_release.py --all # build all artifactsBuild the installable Desktop .mcpb on each target platform with an
approved runtime environment:
python scripts/build_release.py --desktop --runtime-root <venv>The builder is deterministic: a fixed DOS timestamp is stamped on every zip entry so the artifact hashes are stable across rebuilds.
Proprietary — Copyright © 2026 Guillaume Lessard / iD01t Productions. See LICENSE.md.
The underlying qector-decoder-v3 backend is separately licensed:
- ✅ Free for personal, academic, educational, and non-commercial research.
- 💼 Commercial use (company R&D, SaaS, hosted API, OEM, redistribution) requires a paid license — see qector.store/pricing.
- 🔄 60-day commercial evaluation, creditable against a license.
| Website | www.qector.store |
| Commercial licensing | admin@qector.store |
| Support | admin@qector.store |
| Pricing | qector.store/pricing |
| Security disclosure | admin@qector.store (private) |
QECTOR Claude Plugin v1.0.4
Built on qector-decoder-v3 v1.0.0 (Rust/PyO3 core)
© 2026 Guillaume Lessard / iD01t Productions