Webflow’s early 2025 move into the Model Context Protocol (MCP) space revealed a significant gap between traditional software design and the needs of AI agents. The company initially tried wrapping existing developer APIs for use by autonomous systems. This approach failed quickly. APIs designed for human developers assume the user can read documentation, manage state, and manually recover from errors. Agents operate with far less context, relying entirely on the API surface to guide their actions.
Direct exposure of standard APIs created tools that were too low-level. Simple tasks required excessive “chatty” interactions, leading to high latency and subtle failures. The team realized that agent APIs must prioritize execution reliability over flexibility. They redesigned tools around user intents rather than specific implementation details. A request to update a homepage hero section, for instance, became a single, declarative call rather than a chain of low-level operations.
Shifting to intent-driven design
This shift indicates that the industry is moving away from RESTful patterns as the primary interface for software. Future platforms may need to be built as programmable environments from the ground up, where code and state replace remote procedure calls as the standard method of interaction.
Related: DeepSeek smaller model beats flagship version
However, scaling this approach presented a new challenge known as “tool explosion.” With too many specific workflows, the tool surface became difficult for agents to search. Webflow solved this by grouping capabilities into layered domain tools and introducing filesystem abstractions. This allows agents to manipulate projects as code, leveraging the model’s strength in reasoning over structured text.
By shifting the interaction model from API orchestration to structured state manipulation, the system reduces the burden on the model. Instead of discovering and selecting from hundreds of endpoints, an agent can operate within a familiar programming environment.
Infrastructure and observability
The infrastructure for these systems differs significantly from standard web services. Webflow built its server on Cloudflare Durable Objects to support stateful sessions. This architecture preserves context across multiple calls without forcing the model to reconstruct its state repeatedly. The system also routes operations between headless APIs and a live browser canvas via a WebSocket bridge.
Understanding agent behavior required new observability tools beyond traditional logging. Standard server logs could not explain why sessions failed despite no technical errors. The team implemented a layer called MCPCat to track intent and missing-tool attempts. This analysis revealed that 27.7% of sessions involved direct design on the canvas, a figure that surprised the team. Additionally, 58.7% of usage focused on just three core workflows.
Related: Small firms turn to AI for customer help
Missing-tool attempts exposed capabilities that agents expected but could not find. Session replay showed workflows where the model repeatedly retried or changed direction without producing a hard error. The system often appeared healthy from an infrastructure perspective while the agent experience was failing silently.
Scaling adoption
Growth accelerated after the launch of a hosted MCP connector in early 2026. Session volume grew by 6.7X in the three months following the release. The company is now focusing on “skills” that encode product judgment and domain conventions. This evolution suggests that successful integration requires more than just API access; it demands an environment where agents can apply complex reasoning to achieve specific outcomes.
As adoption increases, the challenges are shifting from protocol mechanics to product experience. The focus is expanding to include richer site context and a deeper skill layer. The goal is to ensure that agent actions actually help users accomplish meaningful work rather than just executing isolated commands.
