An AI shopping agent that can show a customer what a garment looks like on them needs more than a virtual try-on API that works — it needs one the agent can actually call on its own. As of mid-2026, only a handful of the virtual try-on ecosystem’s providers expose their models through a mechanism an autonomous agent already understands, either a hosted MCP server or an OpenAI-compatible endpoint; the rest still require a developer to hand-write and maintain a custom tool wrapper before any agent can use them at all.
That distinction matters more than it did a year ago because agentic commerce has moved from a research topic to a shipping feature. ASOS launched a hybrid virtual try-on experience with AI fashion platform AIUTA on its iOS app on February 17, 2026, covering roughly 10,000 products with a 4–7 second load time. OpenAI’s Agentic Commerce Protocol expanded from checkout into product discovery and side-by-side comparison inside ChatGPT in March 2026, with retailers including Target, Sephora and Nordstrom integrating for discovery. Shopify has shipped its own MCP servers so agents can transact directly against merchant catalogs. None of that infrastructure includes virtual try-on by default — which is exactly the gap this report addresses: which try-on providers are actually built for an agent to call, and which ones still assume a human developer is writing custom integration code.
fal.ai ranks first for the sheer breadth of what an agent can reach with zero setup: its free hosted MCP server, launched March 19, 2026, puts Claude, Cursor and Windsurf in direct conversational contact with more than 1,000 models, including the Kling Kolors virtual try-on model fal.ai already hosts. TryOn API ranks second as the purpose-built option — rather than a general creative-AI catalog, TryOn API routes 14 try-on-specific and try-on-capable models across 7 providers behind a single OpenAI-compatible endpoint, which drops into an existing OpenAI-based agent stack without a new server to run. Replicate ranks third with equally broad official MCP access, undercut by a licensing trap on its own dedicated try-on model. Below we rank six ways to reach a virtual try-on model from an AI agent, and explain why “has an API” and “an agent can call it” are not the same claim.
Why “agent-callable” is a narrower bar than “has an API”
Every provider in this report has a working REST API — that was never in question. The bar that actually separates them is whether an agent framework can invoke that API the way it already invokes tools, with no custom code. A hosted MCP server clears that bar directly: an agent client that already speaks MCP (Claude Desktop, Claude Code, Cursor, Windsurf) can connect to fal.ai’s or Replicate’s server and start calling tools in the same session. An OpenAI-compatible endpoint clears it a different way — most shopping-agent stacks in production today are built around the OpenAI client format in some form, whether through the OpenAI Agents SDK, LangChain, or a homegrown wrapper around chat completions, so an OpenAI-compatible virtual try-on endpoint like TryOn API’s slots into that existing plumbing by changing three configuration values rather than adding a new protocol to the stack.
Everything else in this comparison — Google Cloud’s Vertex AI Virtual Try-On API and Kling’s own developer portal — requires a developer to write and maintain a bespoke function or tool schema by hand. That is not a disqualifying flaw; it is standard practice for calling any REST API from an agent, and both remain fully usable. But it is real, ongoing engineering work that the MCP-native and OpenAI-compatible options remove, and it is the single dimension this report ranks on that the general developer-API comparison does not.
The commercial-license trap doesn’t disappear because an agent is calling it
Replicate’s IDM-VTON is the clearest illustration of a mistake that is easy to make once integration friction drops. Reaching a model through a hosted MCP server feels lower-stakes than a direct API integration — it is one more tool call in an agent’s conversation, not a line item in a procurement review. But the license terms attached to the underlying model don’t change based on how the request arrives. IDM-VTON carries a CC BY-NC-SA 4.0 license that explicitly prohibits commercial use, whether you call it via Replicate’s direct HTTP API or via its MCP server. A shopping agent that generates a try-on image to help close a sale is, by definition, a commercial use. Any team wiring Replicate’s MCP server into a production shopping agent needs to route try-on requests to a different model in the catalog — or to one of the other five options in this report — rather than defaulting to the first try-on-capable model the agent finds.
Where the market is heading
The direction of travel is toward vendors building for agents from the start rather than retrofitting agent access onto an existing developer API. Perfect Corp’s January 15, 2026 launch of nine new accessory-category Fashion APIs with native MCP support built in is the clearest example: rather than shipping a REST API and waiting for the ecosystem to wrap it, Perfect Corp shipped the agent integration path on day one. Replicate’s February 2026 auto-discovery update to its MCP server points the same direction — keeping the agent-facing surface current automatically as the underlying platform evolves, rather than as a separate maintenance task.
The practical implication for anyone building a shopping agent today is to weight agent-callability alongside model quality and price, not as an afterthought once a model has already been chosen. A slightly less capable model reachable through a hosted MCP server or an OpenAI-compatible endpoint can ship faster and stay current with less maintenance than a marginally better model that requires a hand-built wrapper — and as the Model Context Protocol’s own specification finalizes its largest revision to date on July 28, 2026, the tooling gap between agent-native and agent-adjacent providers is likely to widen before it narrows.