QECTOR Claude Plugin

Verified quantum error correction for Claude. Local, deterministic, and mathematically guaranteed on every machine you work on.

QECTOR integrates the published qector-decoder-v3==1.0.0 engine with Claude Code and Claude Desktop, turning both into a production grade QEC engineering environment. Every correction returned by an MCP server is verified against the parity relation H c = s (mod 2) before it leaves the process, every claim carries a machine checkable provenance class, and the default configuration performs no network request of any kind.

4 MCP servers · 8 stable / 29 research / 3 admin tools · 11 commands · 5 agents · 28 skills · zero egress by default

Latest release — v1.0.6 · Claude.ai marketplace compliance (launchers to scripts/, canonical manifests), environment-agnostic /qec-setup, deterministic artifacts. See the release announcement.

Why v1.0.6

Version 1.0.6 completes the portability and compliance story that 1.0.5 started, then layers on marketplace correctness and universal setup. Two problems drove it: modern operating systems no longer hand you Python on a predictable command name (macOS, Debian, Ubuntu, and Fedora ship python3 without bare python, while Windows adds its own resolver quirks), and the claude.ai marketplace approval pipeline rejects plugins that ship executables in bin/. This release retires both failure classes, permanently.

1. A universal, compliant launcher ships inside every artifact.

• The plugin archive, the source distribution, and the Desktop bundle all carry scripts/qector-python (POSIX sh) and scripts/qector-python.cmd (Windows). They live in scripts/, not bin/: claude.ai-hosted plugins may not ship bin/ executables because they land on PATH without ever appearing on the admin approval surface. A validator now hard-fails any build that reintroduces bin/.
• Resolution is governed, not guessed. An administrator pinned interpreter wins first, then python3, then python, with the Windows py launcher tried ahead of them all.
• The supported window is enforced, not documented. Only Python 3.9 through 3.13 executes, matching the published native wheel matrix; an out of range machine receives precise remediation guidance instead of a deep import crash.
• Pinning is first class on both surfaces. Set QECTOR_PYTHON to an absolute interpreter path and every launcher honors it above discovery.

2. The marketplace surface is canonical.

Both plugin.json and marketplace.json use only documented fields and a relative same-repo plugin source — the form every official marketplace uses and the only form every sync path can resolve. The Desktop MCPB carries a win32 platform override so Windows always selects the .cmd launcher.

3. Setup works in every environment.

/qec-setup detects sandboxed and remote environments (claude.ai code execution, Cowork, containers without a project checkout) and falls back to native diagnostics instead of dead ending on a missing local script — and it reports honestly which path it used.

Platform support

Local stdio cannot reach web or mobile by architecture; the hosted connector is the engineered path to those surfaces and lands in 1.1.x.

Installation

Prerequisites

Python with spaces in path (e.g. Anthropic Skills and agents) is now handled — the Desktop extension in v1.0.6 uses python directly (no ${__dirname} launcher) and the Code plugin uses scripts/qector-python. If you still see Server disconnected, set QECTOR_PYTHON (see Troubleshooting).

Option A — Claude Code (marketplace, 30 s)

# 1. Add the marketplace (once)
claude plugin marketplace add GuillaumeLessard/qector-claude-plugin

# 2. Install the plugin
claude plugin install qector@qector-tools
# When prompted for Python interpreter: leave empty for auto-resolve,
# or paste an absolute path like C:\Program Files\Python311\python.exe

# 3. Verify
claude plugin list                          # → qector@qector-tools 1.0.6 enabled
claude plugin validate . --strict           # → ✔ Validation passed
python scripts/qector_runtime_check.py      # → status: ok, mcp 1.29.0

Option B — Claude Desktop (single click)

Via UI (recommended):

1. Download qector-claude-desktop-1.0.6.mcpb from the release page.
2. Claude Desktop → Settings → Extensions → Advanced settings → Install Extension → select the .mcpb → Enable.
3. Fully quit Claude Desktop (system tray → Quit) and reopen — the extension needs a full restart to load the 8-tool safe profile.

Via script (same result, auditable):

python scripts/configure_claude_desktop.py --check-only   # preview, no changes
python scripts/configure_claude_desktop.py --confirm      # writes claude_desktop_config.json + installs extension
# Restart Desktop afterwards

Verify Desktop:

• C:\Users\<you>\AppData\Roaming\Claude\claude_desktop_config.json contains qector-library (if you used the script) and extensions-installations.json contains ant.dir.gh.guillaumelessard.qector.
• C:\Users\<you>\AppData\Local\Claude\logs\mcp.log shows QECTOR Server started and connected → initialize → tools/list → result (no Server disconnected).

Option C — From source (auditable, offline)

git clone https://github.com/GuillaumeLessard/qector-claude-plugin.git
cd qector-claude-plugin

# Install for every Python you use (Windows has both 3.11 and 3.12 via py launcher)
python -m pip install -r requirements.txt          # → numpy 2.2.6, mcp 1.29.0, qector 1.0.0
py -3 -m pip install -r requirements.txt           # ← do this too if py -3 exists

# Verify
python scripts/qector_runtime_check.py             # → status: ok, failures: []
python scripts/qector_system_setup.py --check-only # → dry_run_pending_approval
python scripts/qector_system_setup.py --confirm    # → ready, theorem_1_faithful: true

Troubleshooting

The QECTOR surface

Alongside the servers: 11 slash commands for setup, facts, theorems, reproduction, threshold sweeps, Wilson intervals, detector error models, code inspection, benchmarking, sinter templates, and validation; 5 specialized agents (researcher, developer, validator, sysadmin, hardware engineer); and 28 skills spanning math foundations, architecture, BP OSD, two stage CSS, space time codes, decoders deep dive, DEM pipelines, LER methodology, pymatching compatibility, batch decoding, reproducibility, testing strategy, deployment, release engineering, licensing, education, and operations.

Artifacts and integrity

The builder is deterministic: fixed timestamps make every rebuild bit identical, so the hashes below are stable forever.

Read the full README on GitHub →