Quickstart · Tools · Hosts · The Loop · Docs
Verified AIAG core-tool engines — FMEA, SPC, Control Plan and MSA / Gage R&R —
callable by hand in a browser, or by any AI agent through an MCP server and Agent Skills.
In real quality departments, the AIAG / IATF-16949 core tools live in disconnected spreadsheets and one-off apps. A failure mode found in an FMEA never becomes a control on a Control Plan, and an out-of-control point on an SPC chart never makes it back to the FMEA's risk rating. The methodology describes a closed loop; the tooling almost never implements one.
Quality Platform builds that loop for real — credible standalone tools first, everything they share
promoted into a single typed core (quality_core), then wired into an end-to-end workflow you run
on demand over a project's files. The last leg is deliberately not automatic: SPC evidence
proposes an occurrence rating for a human to review, it never rewrites one. Proven on real
semiconductor process data.
 🛡️ FMEA Risk Analyzer — RPN & AIAG-VDA Action Priority, risk-tier triage, auto-generated insight. |
 📈 SPC Control Charts — X̄-R / I-MR / c-charts with Western Electric & Nelson rule overlays. |
 📊 Process Capability — Cp/Cpk/Pp/Ppk with a stability gate: no capability claim on an out-of-control process. |
 🏭 Unified shell — every tool under one st.navigation surface, one theme, one URL.
|
| It is | It is not |
|---|---|
| On-demand analysis over files you supply — every result comes from a call you or your agent makes | Not real-time or 24/7 monitoring. Nothing polls, streams or watches a line |
A set of standards-anchored engines, every constant cited in an ASSUMPTIONS_LOG.md |
Not a data historian and not an EQMS — no database, no equipment integration, no retention or e-signature layer |
| A loop that proposes an occurrence rating from SPC evidence, with an action pointing at the evidence file | Never an autowrite. Cause.occurrence is never overwritten — a human reviews the proposal and decides |
| Local by default — the MCP server's stdio transport runs on your machine, so your data never leaves it | Not a replacement for engineering judgment. The tools compute, cite and refuse; the engineer decides |
Every one of those lines, with the reasoning behind it, is on the docs site: Limitations.
Quality work happens where the engineer already is. So the engines ship as MCP tools and Agent
Skills first, and as a UI second — the same quality_core code underneath both, so the two cannot
drift apart.
- Local, no database, no upload. stdio is the default transport: your agent host launches the server as a local process and talks to it over pipes. There is no port, no token and no service — the data stays on the machine that owns it.
- One config, any host. 49 tools behind a single stdio config block that four named hosts accept nearly verbatim. Which hosts, which quirks, and — honestly — which configurations have actually been run rather than merely written: Hosts.
- The agent does not do the arithmetic. The point of the tool layer is that
fmea_score(8, 5, 6)returns{"rpn": 240, "action_priority": "Medium"}from tested, coverage-gated code — not from a model's head. - Roadmap. MCP + skills today. A standalone status-product website is a Phase-2 item that is not built and not live; nothing on this page depends on it.
Clone first — nothing is published to a package index yet (uvx quality-mcp does not work;
#292 is open).
git clone https://github.com/Siddardth7/quality-platform.git
cd quality-platform
uv syncuv run python -m mcp_app.server # from the workspace rootRegister it — this is the Claude Desktop shape (root key mcpServers); Cursor and Gemini CLI take
the same block in their own config files, and VS Code uses servers instead:
{
"mcpServers": {
"quality-platform": {
"command": "uv",
"args": [
"run",
"--directory", "/absolute/path/to/quality-platform",
"python", "-m", "mcp_app.server"
]
}
}
}Then ask your agent: "Score this failure mode: severity 8, occurrence 5, detection 6" — the host
should call fmea_score and report the tool's result, rpn 240 / "Medium". health() and
version() are the cheaper liveness checks. Run the whole loop with
run_project_loop("examples/secom-quality-loop"), or a capability study with spc_capability.
Per-host config blocks, the quirks table and the PASS/PENDING record: Hosts →
apps/mcp/docs/HOSTS.md.
uv run streamlit run app.py # the whole platform — one URL, every tool
uv run streamlit run apps/spc/app.py # a single app standaloneRun a single app from its own directory
cd apps/fmea && streamlit run app.py # FMEA Risk Analyzer
cd apps/spc && streamlit run app.py # SPC DashboardEach app still runs unchanged from its own directory.
uv run python -c "from mcp_app.server import run_project_loop; \
run_project_loop('examples/secom-quality-loop')"The write-up — including which files in it are real SECOM data and which are illustrative —
is at docs/demo.md and
examples/secom-quality-loop/.
The architectural payoff: the AIAG core-tools loop, wired end to end and run on real data. It runs on demand over a project directory on disk — one call, four arrows — not continuously.
flowchart LR
FMEA["🛡️ FMEA<br/>score S·O·D →<br/>RPN / Action Priority"]
CP["🧩 Control Plan<br/>failure mode → characteristic,<br/>spec, method, sample plan,<br/>recommended chart"]
SPC["📈 SPC<br/>control charts +<br/>capability (Cp/Cpk)"]
MSA["📏 MSA / Gage RR<br/>is the measurement<br/>system even trustworthy?"]
SECOM[("🏭 SECOM<br/>real semiconductor<br/>process data")]
FMEA -->|"high-risk items<br/>become controls"| CP
CP -->|"auto-configures<br/>the chart"| SPC
SPC -->|"proposed occurrence-rating /<br/>CAPA (human reviews)"| FMEA
MSA -.->|"prove the gage<br/>before trusting the chart"| SPC
SECOM -.->|"runs through<br/>every tool"| SPC
classDef live fill:#0b1220,stroke:#e65100,stroke-width:2px,color:#fff;
class FMEA,SPC,CP,MSA,SECOM live;
Two things the diagram cannot say by itself:
spc/results/*.jsonis an input, not an output. No arrow writes it — it is the trace of a prior, separate charting session, read off disk as a precondition. With none there, the feedback leg legally no-ops and returnsnull.- The feedback arrow proposes. It attaches a candidate
Actionwhoseownerpoints at the evidence file and leavesCause.occurrenceuntouched — "it NEVER writes a new rating, it only proposes one for a human to review." Re-running with unchanged inputs rewrites nothing.
Full detail: The loop.
| Tool | What it does | Status |
|---|---|---|
| 🛡️ FMEA Risk Analyzer | Failure Mode & Effects Analysis — RPN + AIAG-VDA Action Priority, editable S/O/D scales, relational model (Function → FM → Effect / Cause / Control), action tracking, Pareto + risk heatmap, Excel/PDF/CSV export |  |
| 📈 SPC Dashboard | Statistical Process Control — variables & attributes control charts, Western Electric / Nelson rules, Cp/Cpk/Pp/Ppk with a stability gate, live disturbance simulator | |
| 🧩 Control Plan connector | Turns FMEA failure modes into a Control Plan (characteristic, spec, method, sample plan, recommended chart) — the APQP-adjacent bridge that closes the loop | |
| 📏 MSA / Gage R&R | Measurement Systems Analysis — Gage R&R (Average-and-Range by default, or ANOVA with the part×appraiser interaction), %EV/%AV/%GRR/%PV vs study & tolerance, ndc, accept/marginal/reject vs AIAG thresholds |
|
| 🔌 MCP server | 49 tools over the same engines — Meta (2) · FMEA (5) · SPC (19) · project-file arrows & the loop (4) · export/report (13) · Control Plan (4) · MSA (1) · private-corpus RAG (1). Full grouped catalog: tool catalog | |
📚 quality-research skill |
Answers "why does the standard say X" by calling qdb_answer_question for a cited, page-located answer, falling back to this repo's public ASSUMPTIONS_LOG.md citations when the endpoint is unreachable. Explicitly not for running an engine — that is the fmea / spc / msa / control-plan skills |
|
| 🏭 SECOM case study | The whole platform run on real semiconductor sensor data — SPC, yield/DPPM, Pareto of failing signals. No Cp/Cpk and no Gage R&R: the dataset structurally supports neither, and both are refused rather than invented |  |
Standards context: FMEA — AIAG-VDA (2019) + AIAG FMEA-4 · SPC — AIAG SPC 4th Ed. · capability target Cpk ≥ 1.33. The AIAG-VDA Action Priority table is verified cell-by-cell against the primary handbook.
Six hosts have a written, schema-correct MCP-server config in
apps/mcp/docs/HOSTS.md. What has actually been run is tracked
separately from what has been written, on purpose:
| Host | Transport | Config verified | Worked example run |
|---|---|---|---|
| Claude Desktop · Cursor · VS Code | stdio | PASS ✓ (shape) | PENDING — GUI apps, not launchable headless |
| Gemini CLI | stdio | PASS ✓ — the host spawned the server and enumerated its tools | PENDING — the host's non-interactive permission gate auto-denied the call |
| Claude.ai · ChatGPT | http | PASS ✓ (shape only) | PENDING — no hosted endpoint and an OAuth-vs-bearer gap (#355) |
| Claude Code · Codex CLI · Cursor · Gemini CLI (skill layer) | stdio via skill scripts | — | PASS ✓ on all four — skills/COMPATIBILITY.md |
A PASS ✓ in the config column is not permission to say the host works. The last row is the
skill-script layer, which is a different layer from native MCP-server registration — the two do
not overlap. HTTP transport is opt-in (MCP_TRANSPORT=http), always authenticated, and fails
closed. The registry manifests (server.json, smithery.yaml, glama.json) exist but none is
live-published.
The public validation story, in three checkable pieces:
- Every constant is cited. Each app carries
docs/ASSUMPTIONS_LOG.mdlisting every AIAG/ISO constant, threshold and quotation with its source. A value cannot change without its log changing. - MSA's citations are machine-checked.
apps/msa/docs/CITATIONS.tsvis a manifest asserted in CI byapps/msa/tests/test_citations.py— a drifted or fabricated quotation fails the build, not a review. - Nine coverage gates at 100% with branch coverage on (see below). Where no published standard exists, the module says so in its own docstring rather than implying one.
Refusals count too: no capability claim on an out-of-control process, no Gage R&R on SECOM (it has no part/appraiser/trial axis), no Cp/Cpk on SECOM (it ships no tolerances). More: Standards & fidelity.
A uv workspace monorepo: four Streamlit apps mounted under one shell, SECOM as an engine-only
member (a tested library, deliberately not mounted), and an MCP server exposing the same
engines as tools. Every cross-cutting concern is written once in quality_core and consumed by
all of them.
flowchart TB
subgraph Agents["🔌 Agent hosts · MCP + Agent Skills"]
MCP["apps/mcp<br/>FastMCP server · 49 tools"]
end
subgraph Shell["🏭 Unified shell · app.py (st.navigation)"]
Home["Landing + one theme + one nav"]
end
subgraph Apps["Apps · mounted in the shell"]
FMEA["🛡️ FMEA<br/>apps/fmea"]
SPC["📈 SPC<br/>apps/spc"]
CP["🧩 Control Plan<br/>apps/controlplan"]
MSA["📏 MSA<br/>apps/msa"]
end
subgraph Engine["Engine-only · library, not mounted"]
SECOM["🏭 SECOM<br/>apps/secom"]:::engine
end
subgraph Core["📦 packages/quality-core → import quality_core"]
Schema["schema/<br/>flat + relational contracts (Pydantic v2)"]
IO["io/<br/>validated ingest · CSV/Excel/PDF export"]
Scoring["scoring.py<br/>RPN · AIAG-VDA Action Priority"]
Theme["theme/<br/>palette · style"]
end
Home --> FMEA & SPC & CP & MSA
MCP --> FMEA & SPC & CP & MSA
FMEA --> Schema & IO & Scoring & Theme
SPC --> IO & Theme
CP --> Schema & IO & Scoring
MSA --> Schema & IO & Theme
SECOM --> IO
SECOM -.reuses the SPC engine.-> SPC
classDef engine opacity:0.85,stroke-dasharray:4 4;
Why it's built this way
- Shared core, consumed many times.
quality_core.ioowns CSV/Excel/PDF export (formula-injection safe) and validated ingest — so upload validation and export are guaranteed identical across tools, and across the UI and the MCP tools. That's the economic argument of a monorepo, made concrete and coverage-gated at 100%. - Schema promoted only when stable. Contracts lived inside the FMEA app until they earned promotion
to
quality_core.schema— deferred extraction, done once, correctly. - History preserved. The FMEA and SPC apps were previously standalone repos, migrated here with full commit history intact — the histories are part of the engineering story.
The whole workspace shares one quality bar (ruff.toml, mypy.ini, pytest config in
pyproject.toml). It runs locally and, identically, in CI on every push and PR to main — a
protected branch that requires the gate to pass before merge.
uv run ruff check . # lint + format check
uv run mypy # strict static types
uv run pytest --cov # 2502 tests + coverage across core + appsCoverage gates — CI-enforced, cannot silently regress:
| Surface | Bar |
|---|---|
quality_core.io — shared export + ingest |
100% |
quality_core.schema — shared FMEA contracts |
100% (line + branch) |
| SPC testable surface — engine + simulation + visualizer + exporter | ≥ 95% |
Workflow discipline: one logical change per commit (conventional commits) · one issue at a time · multi-agent code review before finishing · push → CI green → close issue → tag a release each week · if a week can't ship green, cut scope, not quality.
Twelve tracked weeks, one release each, ending on a portfolio-grade v1.0.0.
| Phase | Weeks | Focus |
|---|---|---|
| A · Foundation | 1–2 | Monorepo, shared core, shell, one CI gate · v0.1–v0.2 ✅ |
| B · Standards-correct cores | 3–5 | AP-native + relational FMEA, shared validation/export · v0.3–v0.5 ✅ |
| C · Integration & core-tool completion | 6–9 | Control Plan → close the loop → MSA → SECOM real-data case study |
| D · Depth & legibility | 10–12 | Modern SPC depth, DOE on SECOM, then a hardening pass → v1.0.0-portfolio |
An explainable AI FMEA copilot (LLM + RAG + eval harness) is a documented, unscheduled future phase. The full plan — vision, diagrams, week-by-week detail — lives in ROADMAP.md.
quality-platform/
├── app.py # unified platform shell (st.navigation)
├── shell/ # landing page + shared chrome
├── mkdocs.yml # docs site config (MkDocs Material)
├── docs/ # docs-site pages + engineering docs
├── skills/ # Agent Skills — fmea · spc · msa · control-plan · project-loop · quality-research
├── examples/ # worked examples — secom-quality-loop/
├── ROADMAP.md # the full project guide (vision, diagrams, 12-week plan)
├── packages/
│ └── quality-core/ # shared core → import quality_core
│ └── src/quality_core/
│ ├── schema/ # flat (FMEARow) + relational (Function→FM→…) contracts
│ ├── io/ # validated ingest · CSV/Excel/PDF export (injection-safe)
│ ├── scoring.py # RPN · AIAG-VDA Action Priority
│ └── theme/ # palette · style
└── apps/
├── fmea/ # FMEA Risk Analyzer (full original history preserved)
├── spc/ # Manufacturing SPC Dashboard (full original history preserved)
├── controlplan/ # Control Plan connector — FMEA → characteristic/spec/method/chart
├── msa/ # MSA / Gage R&R — Average-and-Range (default) + ANOVA
├── secom/ # SECOM real-data case study — engine-only, not mounted in the shell
└── mcp/ # MCP server — the engines as 49 tools (FastMCP, stdio)
Migrated from the standalone repos
fmea-risk-analyzer and
manufacturing-spc-dashboard,
now archived → moved here.
The full docs site — quickstart, per-engine pages, the 49-tool catalog, the host matrix, the worked example, standards and limitations — is at:
→ siddardth7.github.io/quality-platform
| In this repo | What it is |
|---|---|
docs/demo.md |
The SECOM worked example, end to end |
docs/tools.md |
The 49 MCP tools, grouped by method |
docs/limitations.md |
What this platform deliberately does not do |
apps/mcp/docs/HOSTS.md |
Per-host MCP configuration + PASS/PENDING record |
docs/DEFINITION_OF_DONE.md |
The contract every change is held to |
New here? Start with the docs site or the ROADMAP.md · Contributing? See CONTRIBUTING.md and the Definition of Done.
Manufacturing-quality engineering, built like software — typed, tested, and shipped weekly.