Python SDK
The primary surface for process engineers and data teams: heats, molds, pours, plans, predictions, windows and audit queries as first-class objects.
Build on thefoundry loop.
A Python and REST surface over the agent loop, OPC UA and MQTT at the edge, DICONDE ingest for inspection media, and a plan object that carries its own process window, confidence and explanation.
12+Connector families
41msp95 edge inference target
100%Actions logged
A plan, a window,an approver.
Everything an agent proposes is a plan object. A plan knows its process window, its confidence and how to explain itself — and applying it requires an approver identity.
- `plan.within_window` is checked before anything can be applied
- `plan.explain()` returns the grounded reasoning and its citations
- `plant.apply(...)` requires an approver and writes to the audit log
# Ask the pour agent for a bounded setpoint move on cell P-04
from smelteon import Plant
plant = Plant("plant-04", edge="foundry-edge-02")
heat = plant.heats.current(line="ductile-iron-1")
plan = plant.agents.pour.plan(
heat = heat.id,
part = "crankshaft-4cyl-rev-c",
mold_state = plant.agents.mold.state(),
objective = "minimise shrinkage porosity at journal 3",
twin = "castwin:solidification-v7",
)
# Every move is checked against the approved process window first
if plan.within_window and plan.confidence > 0.94:
plant.apply(plan, mode="bounded", approver="metallurgist-on-shift")
else:
plant.escalate(plan, reason=plan.explain())$ smelteon edge install --site plant-04 --cell P-04detecting hardware ......... Jetson Thor (edge)installing runtime ......... triton + tensorrtregistering connectors ..... opcua://plc-p04, robot://fanuc-r30ibautonomy ................... shadow (default)$ smelteon models list --cell P-04pourbot-fill-v6 edge serving golden 0.951poreon-ct-v9 edge serving golden 0.964meltra-chem-v4 edge serving golden 0.938castwin-solid-v7 cloud on-demand validated$ smelteon autonomy promote --cell P-04 --to assistgolden-set gate ............ pass (0.951 ≥ 0.940)process window review ...... signed: metallurgist-on-shiftsafety review .............. required for bounded, not assistautonomy ................... assist
Operate it fromthe shift office.
The CLI speaks in heats, molds, pours, windows and approvals — the vocabulary the plant already uses.
Four ways in.
REST & webhooks
Read the loop, subscribe to gradings, non-conformances and escalations, and push context from your MES or ERP.
OT protocols
OPC UA, MQTT and Modbus at the edge for furnace, molding, pouring and robot integration, with historian ingest for the process record.
DICONDE ingest
X-ray and CT volumes, ADR results, dimensional scans and vision streams ingested as inspection ground truth.
What actually runsinside your plant.
-
Cell node
Jetson Orin or Thor beside the PLC and robot controllers, running TensorRT and Holoscan inference for vision, pour and defect models with no inbound cloud dependency.
-
Factory server
Triton Inference Server with the model router, RAPIDS ETL for historian data, and the local audit buffer that replays to cloud on reconnect.
-
Model router
Routes each step to the best and cheapest model: fine-tuned open models for high-volume steps, frontier models for metallurgy and safety reasoning.
-
Twin bridge
Requests Castwin simulations for new geometry, alloy or process changes and receives promoted setpoints back into the cell plan.
-
Connector layer
Read-first connectors to furnace, molding, pouring, inspection, robot and MES systems; write-back only through the approved-window gate.
-
Audit sink
Append-only local log of every plan, approval, action, override, escalation and grading, replicated to the tenant's audit store.
What we speak,and how.
| System | Protocol | Direction |
|---|---|---|
| Induction & holding furnaces | OPC UA / vendor API | Read, then bounded write |
| Spectrometry & thermal analysis | File drop / API | Read |
| Molding & coremaking lines | OPC UA / PLC tags | Read, then bounded write |
| Auto-pour & ladle systems | Vendor API / PLC | Read, then bounded write |
| HPDC / LPDC cells | OPC UA / vendor API | Read, then bounded write |
| Robot controllers (KUKA, FANUC, ABB) | Vendor API / safety-rated I/O | Read, then bounded motion |
| X-ray, CT, vision, dimensional | DICONDE / vendor SDK | Read |
| Foundry MES & ERP | REST / database | Read and write records |
| Historians | OPC UA / SQL / time-series API | Read |
Every integration ships with a documented failure mode. Write-back is enabled per connector only after a safety review.
How we buildfor a plant floor.
Local first
The loop runs at the cell. Cloud is for training, validation and fleet updates — never a dependency for control.
Fail to observation
Degraded modes fall back to sensing and prediction, never to uncontrolled action. A confused agent watches; it does not experiment.
Everything explains itself
Plans carry their grounding, citations, confidence and window. If it cannot be explained to a metallurgist, it does not ship.
Promotion is a gate, not a deploy
Model and prompt changes pass golden sets, LLM-as-judge evaluation and twin scenario tests in CI before they reach a plant.
8–24
Camera or detector streams per line
30–120
FPS vision inference per line
12–40
Models served per factory
< 60s
Line reschedule target
[ASPIRATIONAL] Engineering targets from Smelteon's GPU utilization plan.
Where to go next.
Docs · deployment
Deploy the edge runtime
Hardware requirements, network posture, connector setup and the shadow-mode checklist.
Docs · concepts
Model the loop for your plant
Heats, molds, pours, plans, windows, gradings and genealogy — the object model behind the API.
Docs · autonomy
Promote autonomy safely
Golden-set gates, process-window authoring, safety review and rollback configuration.
Security
Security architecture
Isolation, alloy IP protection, grounding, evaluation and the immutable audit log.
Things engineersask on day one.
Python is the primary SDK, with a REST surface and webhooks for everything else. At the edge, integration is over OPC UA, MQTT, Modbus and vendor APIs.
Yes, on Factory and Enterprise tiers. Custom models are served through the same Triton-based router and are subject to the same golden-set and promotion gates as ours.
Staged promotion with instant rollback. A new version serves shadow traffic alongside the incumbent until its golden-set performance and twin scenario tests support promotion.
Read the docs.Then walk the line.
The fastest way to understand the loop is to see it against your own cell and your own inspection data.