solidSF open-sourced solidcv-cad, which takes video in and puts out STEP, GLB or IGES files for single parts and full assemblies. A sister repo pulls a simulation model from real life. Reverse engineering a worn part could start with a phone video.
The skills cover IEC 61131-3 PLC work, preliminary IEC 62443 OT security assessments, paperwork for authorized OT testing and IEC 61000 EMC planning. They run locally with no network access, and the author says plainly that output is guidance, not compliance.
For controls and OT security engineers: a first draft of the review and gap assessment work, to be checked by a qualified engineer.
It works on isolated copies of the project and ties every change to the exact PLC model, source hash and revision, with a record of each operation. It also covers DVP serial monitoring.
For controls engineers on Delta hardware: let an agent analyze and edit a project copy while keeping a traceable record of what it changed.
Agility and battery partner Nyobolt say Digit 5 runs 90 minutes and recharges in 9, up from a 2 to 1 run-to-charge ratio on Digit 4. Early access is planned for the first half of 2027 for manufacturing and warehouse operators, and the figures are company-reported.
Search returns web results from DeepSeek's own search engine and Fetch returns a page as Markdown, each at $0.75 per 1,000 successful calls. Calls through DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users until billing starts Oct 31 at 16:00 UTC.
For teams building agents that watch supplier sites, prices or regulations: a cheap way to give them live web lookups.
Walking up to a cabinet with no register map is risky, because a wrong poll can fault live equipment. This tool stays in listen-only mode, infers baud rate, node IDs and register maps from traffic already on the wire, and hands the results to an agent.
For controls engineers doing brownfield surveys: map an undocumented Modbus, CAN or J1939 network before anyone touches it.
It gives an agent the robot's topics, maps and camera feeds, and a read-only mode removes every tool that can send a command. Look first, touch later is the right order for agents near moving equipment.
For automation engineers running inspection robots, crawlers or AMRs: let the agent read the robot's state before you ever let it publish a command.
Given only the model number, Claude Code found the TV on wifi, removed 25 Xiaomi packages and wrote a new launcher. The same move applies to locked-down devices on a plant network: an agent can find and reshape them fast, which cuts both ways for OT security.
A prototype walks kids through how a pump works, step by step. The same format could teach new hires how the equipment in front of them works.
The alert was never the bottleneck; finding the shutdown window and the spare part was.
In most plant design tools the P&ID and the 3D model are separate datasets and mismatches turn up late in a validation check. This prototype keeps permanent IDs apart from editable tags, so renaming or moving a pump updates every output or flags it as pending.
For piping designers and package engineers: see what a single source of truth for a skid could look like before the next revision fight.
Describe the data format once and it draws plots, gauges and maps, records sessions and exports them. The free build covers serial, Bluetooth and network sockets; Modbus, OPC UA, MQTT, CAN and Siemens S7 need the paid Pro build.
For instrumentation techs and automation engineers: a quick way to see what a device is actually sending before anyone builds on that data.
NVIDIA Research treats training as skill refinement instead of gradient descent, so the result is a repo of robot skills rather than model weights. The loop ships as skills for Claude Code and Codex.
For robotics teams: build a skill library you can read and reuse instead of reprogramming each cell from scratch.
Upload a LiDAR or SLAM scan from any scanner, up to 200 GB per file, and get back a twin with every object separated and editable and robots such as autonomous forklifts, pallet AMRs and palletizing arms placed on your floor. You can try layouts, fleet sizes and routes before a robot arrives.
For plant, warehouse and yard managers planning automation: test the robot layout on your real floor before you buy.
Ask in plain words for selection sets, a hard clash run between mechanical and structural, and only the new clashes.
For VDC and BIM coordinators: skip the clicking and get the clash run you asked for.
His skill wasn't code, it was running the agent like a job: break the work into tasks, set limits, and demand tests and proof. People in the trades already manage work that way.
Boris Cherny of Anthropic says Claude Slides is now how he makes slide decks. It's a quick win for anyone building monthly ops reviews or bid decks.
Anthropic highlighted a wildfire prototype that maps a fire's spread over time and tracks hot spots. Utilities and pipeline operators watch exactly this kind of risk near their lines.
Reason Network is a free, open-source, peer-to-peer engine for connecting agents like Grok Bot, Claude Code and Codex, so teammates' agents don't collide. Teams running several agents need this kind of plumbing.
It exposes a daily index for 300+ markets, each with a BUILD, CAUTION or AVOID verdict, to AI agents. The scores are the vendor's own, and some data needs a free key or a paid plan.
For industrial sales and site selection teams: use it as a first screen for which markets have power, then confirm with the utility.
A gateway that exposes an OPC UA address space as standard REST endpoints with live streams, and as MCP when enabled. It is AGPL-3.0, and the app itself does not enforce HTTP authentication.
For controls engineers: a way to hand plant OPC UA data to an agent or app without deep OPC UA knowledge. Check the license before embedding it, and put authentication in front of it.
Zhaofeng and Leju report about 1,000 hours over two months across 15 stations with a 98% loading success rate, but a robot cycle takes under 40 seconds against about 15 for a practiced worker. The figures are company-reported, and the two companies run a joint venture.
A read-first MCP data tap for OPC UA, Modbus, S7, EtherNet/IP, BACnet and more. It returns a downtime root-cause verdict with cited evidence and audits every call. The analysis is deterministic, and an LLM is optional.
For controls and operations teams: let an agent read plant data with an audit trail. Six tools can send data off the box by design, and one setting turns them off.
Its tools compute properties for 29 fluids, pipe and duct sizing, and liquid, gas and steam networks from reference equations, so the figure doesn't come from the model's guess. The accuracy claims are vendor-reported.
For process and mechanical engineers: each answer comes with its method and validity range, so it can be checked.
Schneider reports third quarter revenue up 17% organic to a record 11.5 billion euros, and says data center demand grew by triple digits. It raised its 2026 target and now expects 12% to 14% organic revenue growth this year. These are company-reported results.
A simulated abrasives plant feeds live OPC UA values to an MCP server that calculates every KPI and traffic light the same way each time, and Claude only picks the tools and writes the answer. It is read only, scopes data by role and logs every tool call.
For plant managers and shift leads: a clean model of how to let AI explain the dashboard without letting it make up the figures.
The authors say real-time AI recommendations that affect the well must work inside a management-of-change framework, with a qualified engineer able to review and override each decision and a full record of actions. Their line: "not autonomous drilling but collaborative intelligence."
Writes stay off unless the server is started with them enabled, and a write is sent once and never retried, so a lost reply comes back as "may have been applied, check first." That is how a tool that can touch a live register should behave.
For controls and instrumentation techs: let an agent pull register values and device IDs for troubleshooting without giving it a way to change anything.
A diagram kept in a text file can be changed by an agent and reviewed line by line like any other change. It maps to CIM, IFC and IEC 61850.
For electrical designers and panel builders: describe sources, breakers, cables and loads in one file and get drawings plus rule checks.
The win came from one small recommendation, not autonomy. Oxy says the Houston center now supports all 12 of its Delaware Basin rigs and a Gulf of America drillship and has made more than 1,000 recommendations.
One user had an AI agent find his phone number on listing sites and file the removal requests, then shared it as a free skill. It's small, but it shows an agent doing a tedious paperwork chore from start to finish.
The post describes Claude Code turning a bracket drawing into an editable model with named dimensions. It flagged a missing wall thickness instead of guessing. 34 of 40 sample parts passed the checks. For engineering teams, the useful part is the flag, not the speed.
Tesla developer accounts include $10 a month in API calls at about $0.002 each, enough for one user to have an agent alert him when his truck is low and not plugged in. Fleet managers could build the same alerts cheaply.
MicroFactory switched a customer robot deployment to frontier models for reasoning, with fast vision code written by the model doing the execution. When something new comes up, the large model steps in and adapts the pipeline.
New browser profiles let an agent scrape pages that need a login. You sign in once, and the session is reused without storing the password. That puts supplier portals and internal sites within an agent's reach, so IT will want a say.
Every scanner, wafer pod and overhead hoist was modeled in Blender and runs in three.js like a strategy game. It's a glimpse of cheap, walkable plant models for layout and training.
Jeff Teper says persistent agents in Copilot Studio can each have an account, so people can email them, add them to Teams or mention them in Word, Excel and PowerPoint. The first features are in preview now. For most industrial companies on Microsoft 365, this is where agents will show up first.
Jerry Liu released ExtractBench Pro, a benchmark for pulling data from long, messy documents with tables, charts and handwriting. Spec sheets, MTRs and turnover packages are exactly that kind of document.
A vision model finds the symbols and plain image processing finds the connections, then it exports CSV, SQL, GraphML and Cypher. The author tested it on one dense open dataset sheet, not on the workflows he lists.
For process and chemical engineers: once a drawing is a table, you can ask which valves sit between a pump and a tank, or what a valve isolates, then check the answer against the sheet.
AI vision reads the symbols and tags, code traces the lines, and the output is scored against public answer keys. The repo says every extracted item is unverified until reviewed, and that no operations review has been done yet.
For controls and reliability teams: it targets the first step of a digital twin project, retyping tag lists by hand, and it ranks items by review risk so engineers know what to check first.
An open-source MCP server for Revit 2022 to 2027 with typed tools for common Revit work, running on the user's machine. The README lists 233 tools.
For capital project teams: a way to let an assistant read and change a Revit model without sending the model off the machine. Test on a copy first.
An SPE series piece says real-time AI recommendations that affect the well should sit inside management of change, with a qualified engineer who can review and override each decision. Its line: the future is collaborative intelligence, not autonomous drilling. Five of the six authors work at Halliburton, which sells drilling software.
Google's Gemini account says you can upload a photo with rough measurements and have Canvas in Gemini build an interactive 3D CAD studio, with live sliders and printable STL export. The example is furniture accessories, and the post names the model as "Gemini 4 Argon". Plant teams could try the same photo-to-part idea for small brackets, covers and fixtures.
Vikas Yadav replied with a link to Google's guide for configuring the Google Workspace MCP servers. The person he answered had asked how to connect Google Search Console to Claude through MCP, and Search Console is not among the products the guide covers. The guide covers Gmail, Drive, Docs, Sheets, Slides, Calendar, Chat, People and Classroom, and it is a Developer Preview. It matters because MCP is how many AI tools now reach company documents, mail and calendars.
Chris Yoo says his team launched a way to pick a model, pick a robot and a task, and run it with real-time inference over the cloud, no local GPU. It's a vendor launch post, so treat it as the company's claim. Robot teams would need to check latency before trusting a cloud link on the floor.
Movez says adding the /anatomy-reveal skill lets Claude make real-size 3D models from a photo or name, exploded views, and motion-blurred reveal films from 1:1 to 16:9. It builds on the /anatomy skill. Product and sales teams could use it for equipment explainers.
For product and sales teams: check the output against the real part before it goes in front of a customer.
The assistant works on a staging sheet and proposes changes with reasons, but it can't write to Maximo. A person presses commit, one record goes first as a check, and an Excel diff report is saved. This is a vendor product under the BSL 1.1 license, free for test environments.
For maintenance and reliability teams: clean up asset and work order data in bulk without giving an AI write access.
A browser tool that models, solves and reports linear static 3D frames, MIT licensed. The author's example shows an agent designing a timber roof from IFC through to report and drawings, and he also sells a course on building such agents.
For project and structural teams: a place to test agent-run frame checks on simple structures.
Mistral released Timestral, its first family of time-series foundation models, in four sizes from 5M to 209M parameters. Mistral says the largest ranks first among pretrained zero-shot models on the GIFT-Eval, TIME and fev-bench benchmarks, and that it runs on a standard server CPU without a GPU. Its own example is forecasting plant sensors such as pump vibration, temperature and pressure, so the claims are worth testing on your own data.
It pulls Modbus, BACnet, OPC UA, S7 and other plant protocols at the edge, then exposes that data to an AI assistant over MCP while publishing northbound on MQTT, Sparkplug B and OPC UA.
For controls: stand up a single binary on a gateway PC, register tags, and let Claude or another MCP client read live OT data without a custom middleware project.