EarthCam has launched construction AI connectors and seven new AI models, putting its jobsite camera analysis inside general-purpose AI assistants. The release, dated 20 October from Upper Saddle River, New Jersey, says the connectors are built on the Model Context Protocol (MCP) and are available “as an app in OpenAI’s ChatGPT and as a connector in Anthropic’s Claude.”
Worth being precise about the rest of that list: the release says the connectors “can be added to other MCP-compatible AI platforms such as Google’s Gemini, Microsoft’s Copilot and Perplexity.” Those are described as compatible destinations, not as shipped integrations. EarthCam’s agentic AI page shows logos for ChatGPT, Claude, Copilot, Gemini, Make, Perplexity and Zapier.
The Architecture Is the Announcement
Strip out the model names and what EarthCam has shipped is a position on a design question every vendor opening itself to agents now has to answer: when an AI agent asks about a jobsite, what does it actually receive?
EarthCam’s answer is that it receives conclusions. The release says the company’s “proprietary AI analyzes jobsite imagery on EarthCam cameras and servers ten times faster than customers’ AI agents interpret raw images,” that customers “receive finished detections, insights and reports that are ready to act on,” and that “in some cases analysis is performed at the edge onboard the camera.” Answers arrive faster, it argues, “because the analysis is complete before a question is asked,” and “AI usage costs drop because agents work from EarthCam’s results rather than processing images themselves.”
That is a real trade and worth naming as one. Three things follow from it, and this is our analysis rather than anything the release claims:
- Cost and latency move to the vendor. A general-purpose agent chewing through gigapixel panoramas burns tokens on every question. Pre-computed detections are cheap to query and arrive as text.
- The vendor’s model defines the vocabulary. An agent can only ask about what EarthCam has already decided to detect. Trench incursions, pallets and dump-truck movements are in scope because there is a model for them; anything outside the catalogue is not a question the connector can answer. EarthCam does say custom analytics are available on request from its data science team, which makes the catalogue extensible — but on the vendor’s timetable, not the agent’s.
- Auditability is the counterweight, and EarthCam addresses it directly: “Every answer can be checked against high-resolution images, timestamps and metadata.” The imagery is not withheld — it is simply not what the agent reasons over.
Contrast that with the other MCP announcements of the past few weeks. When Graphisoft shipped an MCP in Archicad 30 beta and when a Sydney builder’s software opened itself to agents while keeping a human in front of the money, the thing being exposed was a system of record — a model, a ledger — where the data is already structured and the hard question is permissions. Arcadis, in plugging its own MCP server into Autodesk Assistant, was arguing about who holds the keys. EarthCam’s source material is pixels, which are not structured at all, so the question it has to answer first is who does the inference. It has decided: not your agent.
That is also the argument OpenSpace made in September when it pitched the agent era with “AI agents need eyes”. Two reality-capture vendors reaching for the same argument within a month suggests visual data is being repositioned as an input to agents rather than a dashboard for humans. That is our reading of two announcements, not a measured trend.
The Seven Models
The release names all seven, five with descriptions:
- AI Trench Safety Analytics — recognises workers and equipment entering and leaving defined excavation areas, building “a searchable visual history for safety review and incident investigation.”
- AI Housekeeping Hazard Detection — identifies visible debris, clutter and trip hazards as site conditions change, feeding EarthCam’s Jobsite Health analytics.
- Thermal AI Analytics — combines thermometric imaging with object recognition to identify people and activity “in complete darkness and other difficult viewing conditions.”
- AI Delivery Verification — recognises boxes and pallets arriving on site with time-stamped evidence, complementing EarthCam’s existing AI Material Analysis.
- AI Dump Truck Analysis — counts truck movements, analyses load status and distinguishes individual vehicles.
The remaining two, Visual Occupancy Analysis and Retail Signage Detection, are named without further description. The release also cites “expanded schedule automation that compares planned milestones with observed site conditions.”
Four of the five described models are safety or logistics functions, and the pattern is consistent: these are countable events at a fixed camera position. Trench entries, pallet arrivals and truck movements are things a static camera can tally reliably, which is a narrower and more defensible class of claim than progress measurement.
The connectors extend a platform that already existed. Control Center 9, the SaaS environment the detections live in, was introduced on 27 March 2025 according to EarthCam’s own press archive.
What Is the Company’s Claim, and What Is Checkable
The accuracy argument is EarthCam’s own, and we report it as such. The release says that “in EarthCam’s benchmark testing, its proprietary AI identified construction-relevant assets and objects with better context, improved accuracy and no hallucinations — compared with leading general-purpose AI models analyzing the same images.” The stated reason is training data: general-purpose models “learn largely from everyday photos gathered online,” while EarthCam says its AI “has been trained on billions of real-world construction images captured by EarthCam’s own jobsite cameras over three decades.”
Neither the release nor the two EarthCam AI pages we read publishes the underlying benchmark — no model names, no test set, no scores. We looked for it on the company’s own AI page as well as the agentic AI page, and found accuracy described but not measured. So the comparison stands as a vendor benchmark reported without published data, which is how we are leaving it.
On commercial terms the release is specific about availability and silent on price: the connectors are available now as a ChatGPT app and a Claude connector, with custom analytics “available on request from EarthCam’s Data Science team.” What none of the three pages gives is a price, a plan requirement or a regional rollout schedule.
The customer reference is named rather than anonymised. Scott Travis, director of northern operations for GBR Roofing, describes using EarthCam AI to monitor fall-through hazards while replacing rusted roof panels: “We worked with EarthCam to establish exclusion zones, so if someone crossed into one, we would get an alert and could verify that the proper fall protection was in place.”
Brian Cury, EarthCam’s CEO and founder, frames the goal as cross-portfolio querying: “We’re creating a future where customers can ask their AI platform about productivity, safety and jobsite security on any of their projects worldwide and get the answers they need fast.”
Where Contractors Say They Expect AI to Help
The release cites the AGC and NCCER 2026 Workforce Survey, and the survey is worth reading first-hand because its detail cuts in an interesting direction.
The 2026 Workforce Survey drew 1,830 responses. 51% of respondents expect AI and robotics to have a somewhat or very positive impact on construction over the next five years, against 8% expecting a negative impact and 21% calling it mixed; that question drew 775 responses rather than the full panel. Separately, 42% said shortages of their own or their subcontractors’ workers had delayed projects, the most commonly cited cause of delay, down from 45% the year before.
But there is a separate question asking which areas respondents believe AI will have the greatest impact on within their own company over the next five years, with up to three selections and 749 responses. It ranks twelve functions, and the order does not favour this launch. Estimating and bidding leads at 67%, followed by document and contract review (66%), scheduling and planning (62%), project management (49%) and accounting and finance (39%). Safety management sits sixth at 26%, and field operations is last of the twelve at 10%. Those response counts and the full ranking appear in the survey’s national results tables rather than in the analysis document.
Most of what the new models do — trench safety, housekeeping hazards, thermal security, delivery and haulage counts — sits in those lower-ranked bands. Not all of it: the same release mentions expanded schedule automation comparing planned milestones against observed conditions, and scheduling and planning is the third-ranked function at 62%. The delivery and material models also feed project records rather than safety alone, and the survey offers no category that matches delivery logistics exactly — mapping them onto “field operations” is our reading, not the survey’s.
Still, the weight of the launch is on functions contractors rank below the top four, and that is worth stating plainly. Two readings, both ours rather than the survey’s. One is that contractors expect AI to pay in the office and not in the field. The other is that estimating ranks first because it is the function people can already picture being automated, while a camera that flags a trench incursion is unfamiliar enough not to come to mind in a three-pick question. The survey does not adjudicate between them and neither can we.
What EarthCam has that an estimating tool does not is a camera already on the wall. The connector argument is that the inference has been paid for and the index already built, so the marginal cost of asking is a text query. Whether contractors who rank safety management at 26% will go looking for it is the thing this launch will actually test.
