FactorySemantics MES

An open-source, modular, agent-native Manufacturing Execution System.

Run one module on a laptop for a five-person shop, or the full suite across a fleet of plants. Every capability is an API and an agent tool first, a screen second. Every number is honest about where it came from.

Status: pre-alpha, 0.3.0, running simulated plants only (as of 2026-10-12). Orders and routings, OPC UA execution, OEE, quality with SPC and gauges, maintenance, finite-capacity scheduling, serialisation and genealogy, a 95-tool MCP agent surface, an ERPNext connector, and a scoring harness that measures whether the MES told the truth about its plant. New since 0.2.0: the plant packs travel inside the package, so pipx install factorysemantics-mes and then fsmes demo bottling bring up a bottling plant on a machine with no checkout, on Linux and on Windows; an SPC tab with I-MR, X̄/R and p charts where a point can be pressed for the reading, the rules it broke, the gauge, the machine's state and the rest of the line; count data — how many units a lot sent back and what for; maintenance work that reaches the right trade by itself, and a supervisor's screen that reads like a plant; a chat on the AI tab that explains the chart in front of it (it needs ANTHROPIC_API_KEY and costs about $0.12 a question; the floor assistant needs a local Ollama); every floor row carrying the shift it happened on; and a simulated line that plays a whole eight-hour shift with a four-link chain of cause in it — a chiller order nobody started, its outlet climbing, the product reaching the filler colder, fill heights a millimetre and a half low — with no limit broken anywhere, which is the test. The release page is the whole of it. Not yet: a real plant — none has run this, and making a first one possible in shadow mode, beside the MES already in charge, is what shadow mode is for; the guide is written and untried. Also not yet: ERPNext Job Card confirmations — whether Job Card is the right place at all is a question put to the ERPNext community on 2026-09-10 and not yet answered; the thin Frappe app that would list the connector on the Marketplace (decision 0008); and PLC write-back beyond a guarded recommendation queue. See ROADMAP.md for what has landed and what is next, docs/ARCHITECTURE.md for the design, docs/LANDSCAPE.md for why this project should exist at all, docs/operate/compatibility.md for what has been tested against what and when, and docs.factorysemantics.com for the documentation.

Nothing on this page is measured on a real plant. Every number in this repository comes from a simulated line and says so where it appears.

Why this exists

Three facts, verified 2026-08-28 (sources in docs/LANDSCAPE.md):

1. No integrated open-source MES exists. ERP projects (Odoo, ERPNext, Qcadoo) stop at work orders — no machine connectivity, no OEE, no SPC. Data-infrastructure projects (United Manufacturing Hub, Rhize) deliberately refuse to be the MES. SCADA projects have machine data but no manufacturing semantics. Nothing open-source does gauge/calibration management at all. Plants that can't afford a €200k–1.5M commercial MES implementation stitch together Node-RED, Grafana, and spreadsheets — and the ERPNext maintainers themselves published the missing-features list (erpnext#50827).

2. The agent-native lane is open, but the window is closing. Every incumbent (Siemens, Bosch, Rockwell) is bolting AI agents onto closed platforms. Tulip shipped an MCP server; Ignition announced one for 2026. An MES whose core is designed to be operated by agents — permissioned writes, append-only audit, semantic events — with the source open, has no occupant.

3. This isn't starting from zero. The kernel came from a working, tested mini-MES (orders, routings, OPC UA machine layer, OEE, ERPNext round-trip) that was merged into this repository on 2026-08-31. The test plant is a deterministic line simulator that lives in src/fsmes/sim/. See Provenance below.

The north star

Siemens Opcenter-level functional coverage, as open-source modules on one kernel. The table is the target; the marks say how far each item has come as of 2026-10-12 (✅ landed · 🟡 partial · ⬜ not started):

Honesty about scale: Opcenter is thousands of engineer-years. The path here is not cloning it — it's shipping the 20% of each module that covers 80% of a small-to-mid plant, correctly and honestly, module by module, with agents as the force multiplier. Depth grows where users pull.

Principles

1. Kernel + modules. A small kernel (master data, orders, routing, dispatch, execution, events, audit, auth) that is always present; everything else — quality, maintenance, scheduling, gauges — is an optional module. A customer installs only what they need.
2. Standards-shaped, not standards-burdened. ISA-95 nouns for the data model, PackML for machine states, ISO 22400 for KPIs, B2MML only at the ERP border. Never invent a word the industry already has; never implement an XML schema internally.
3. Agent-native. Every function callable via REST and MCP. Read tools are free; command tools are idempotent, support dry_run, and pass approval gates. Agents act on behalf of an identified human, and every action — human or agent — lands in the same append-only audit spine.
4. Never invent production. Counter deltas book production; resets re-baseline and book nothing. OEE reports unknown, never a misleading zero. Unlabelled downtime is reported as unlabelled. Every screen states its data source.
5. Runs anywhere. SQLite on a laptop → Docker Compose in a plant → fleet of nodes with a console. Operator UI is plain HTML/JS with no build step, because a plant PC must render it for years without a node toolchain.
6. Simulation-first. CI runs a full fake plant (deterministic six-station line over a real OPC UA server). Every feature is demoable without hardware. Scripted failures — breakdowns, changeovers, counter resets, micro-stops — are named regression tests.
7. Config, not code, at every plant boundary. Which tags a machine exposes lives in a tag map. What a plant enables lives in a plant pack — one directory, validated offline by fsmes pack check, carrying no code and no secret. Three lab packs exist to prove it: two that disagree about the line, and a third invented to disagree with the pack format itself and to run the same wheel with two modules switched off.

Provenance — where the code comes from

This project is a convergence, and its origins are recorded deliberately:

Clean-room rule: everything here derives from the personal projects above, public standards, and public documentation. Nothing derives from any commercial MES's internals, schemas, or documentation, and no employer data or configuration is ever committed. Domain understanding from running an MES daily informs what to build; proprietary material does not enter this repository.

The FactorySemantics family

• FactorySemantics MES (this repo) — the execution layer: runs the floor, books production, owns the audit trail.
• FactorySemantics — the semantic layer: canonical events and ontology this MES is designed to emit (not yet published; see Provenance).
• factorysemantics.com — the commercial layer: on-prem analytics/ML nodes and the fleet console that consume those events.

The MES makes the data; the semantic layer names it; the intelligence layer learns from it.

Decisions (ratified 2026-08-28)

Getting started

A plant in one command. bottling is the six-station ACME beverage line this project demonstrates: an eight-hour shift replayed over a real OPC UA server, a simulated floor booking lots and naming its stops, and the operator screens reading it.

On Linux or macOS:

pipx install factorysemantics-mes   # or: uv tool install factorysemantics-mes
fsmes demo bottling

On Windows:

pipx install factorysemantics-mes
fsmes demo bottling

It prints where the plant is — http://127.0.0.1:8010/dashboard — and an account to sign in with. fsmes demo bottling stop puts it down; fsmes demo machining is a second plant on 8020. First start copies the pack out of the package and generates the shift from its seed, which takes seconds; after that it starts straight away. Everything it writes — the pack copy, the database, the logs, the generated line data — goes to one per-user directory (%LOCALAPPDATA%\fsmes on Windows, ~/.local/share/fsmes on Linux, ~/Library/Application Support/fsmes on macOS, or wherever FSMES_HOME says). Nothing under the installed package is written to.

fsmes demo with no plant is the older built-in story — one order released against a routing, run down a simulated two-station line over a real OPC UA server, booked honestly, and completed, all in one process.

CI proves both from the artifact rather than the source, on every pull request: the wheel-demo job (.github/scripts/demo_from_wheel.sh) builds the wheel, installs it into a fresh virtual environment in an empty directory and runs the story — and refuses a wheel whose fsmes --version disagrees with its own metadata, or that cannot upgrade a database made by the release before it. The wheel-plant job (.github/scripts/plant_from_wheel.py) does the same for the bottling plant, on Ubuntu and on Windows, and then asks the running plant whether sign-in works, the specifications are there, the OPC port answers, the floor is watching, and a lot has been counted.

From a checkout, on Linux or macOS:

python -m venv .venv && .venv/bin/python -m pip install -e ".[dev]"
.venv/bin/fsmes info
.venv/bin/python -m pytest

On Windows (plant PCs are Windows, and CI proves it):

python -m venv .venv; .venv\Scripts\python -m pip install -e ".[dev]"
.venv\Scripts\fsmes info
.venv\Scripts\python -m pytest

What runs today (2026-10-12, 0.3.0): simulated plants, each one a pack — a single directory that answers which plant is this?, refused offline by fsmes pack check before it touches anything. fsmes plant all start brings them all up; fsmes fleet manages the plants this installation owns and refuses every other one; fsmes fleet console puts them on one page that reads and has no write path. Each plant is a complete MES over a real OPC UA server with its operator screens. fsmes score <plant> replays a scripted hour and reports whether the MES booked what the line produced, whether a planned stop was misbooked as downtime, and whether the scripted breakdown was detected. Two MCP servers: the product’s 95 tools and the simulator’s 13. Ninety-five is every tool-carrying module registered — eleven of the twenty-three modules ship tools — and a pack that switches one off takes its tools away with it; nine of the twenty-three are the kernel and cannot be switched off, fourteen can. The test suite is the gate: CI runs ruff and the whole suite on Ubuntu and Windows, Python 3.12 and 3.13, and the suite again on a real PostgreSQL 16.15, on every pull request.

The public demo is the cutlery plant from labs/cutlery, simulated, serving a copy of one scored hour. The way in is factorysemantics.com/demo, which emails a link; demo.factorysemantics.com is gated and turns away anyone arriving without one. It runs on one desktop. Nothing there is a real factory.

Where things are