Julia Rapczynska
Full Stack Software Developer, Push-Based | Poland
Talk
Stop Teaching Agents to Click: WebMCP and the Agent-Native Web
But why are we teaching machines to reverse-engineer interfaces that we built ourselves?
WebMCP explores a different model. A webpage can expose structured capabilities directly to an agent through document.modelContext: named tools, input schemas and JavaScript execution callbacks. Instead of figuring out which button means “add this session”, an agent can invoke the underlying addSession() operation directly.
In this talk, we’ll look at what WebMCP actually changes in web application architecture, why it is not simply MCP running inside the browser, and how a single domain operation can power both the human UI and the agent-facing interface.
We’ll also look at the less comfortable part: giving an agent structured tools does not make it trustworthy. Authenticated browser sessions, authorization boundaries, tool output and prompt injection still matter.
The interesting shift is not that AI can finally click our websites better.
It’s that clicking may no longer be the web application’s only machine-readable interface.
About Julia Rapczynska
A big part of my current work is exploring how AI can support engineering in a practical, responsible way. I’m interested in how LLM-supported workflows, structured knowledge systems, and internal “second brain” approaches can help developers understand complex codebases, make better decisions, and reduce the friction of working at scale.
I’m an outgoing community person who enjoys turning real engineering challenges into talks, workshops, and discussions. I like sharing not only the final solution, but also the messy path behind it: the trade-offs, failed attempts, migration stories, architecture debates, and lessons that are usually the most useful for other developers.
Workshop
AI at Zero Token Cost - Build Faster, Private, Local-First AI Applications - Master Class (In-Person & Online)
AI is moving into the browser, and that changes the economics of AI-powered applications.
With browser-native AI APIs, tasks such as prompting, summarization, writing, translation, and proofreading can run directly on the user's device. This can reduce cloud inference costs, API dependencies, data transfer, and token usage, potentially to zero for supported workloads.
Combined with MCP and WebMCP, applications can expose structured tools and actions directly to AI agents without relying on expensive, context-heavy DOM automation.
In this hands-on workshop, you will learn how to build local-first, agent-ready web applications designed around one goal: do more AI work in the browser, send less to the cloud, and minimize token costs.
High-Level Workshop Agenda
Block 1: The AI-Native Browser
From Cloud AI to Browser AI
- Tokens, context windows, latency, and API costs
- What changes when inference moves into the browser?
- Local-first, cloud-first, and hybrid AI architectures
- Privacy, latency, offline capabilities, and cost
Browser Built-in AI
- Understanding the browser-native AI architecture
- Models, capabilities, and feature detection
- Downloading and initializing browser models
- Capability detection and progressive enhancement
Coffee Break
Block 2: Prompt API
- Interact with the browser's built-in language model
- Create and manage sessions
- Use system prompts and initial prompts
- Stream responses
- Generate structured output
- Manage context windows
- Explore multimodal possibilities
Lunch Break
Block 3: Task-Specific Browser AI APIs
Discover why specialized APIs can be preferable to generic prompting.
Summarizer API
- Summarize documents, articles, and application content
- Configure summary types, formats, and lengths
- Stream generated summaries
- Summarize large documents with small context windows
Coffee Break
Block 4: Model Context Protocol and WebMCP
Model Context Protocol (MCP)
- Understand why agents need tools instead of prompts alone
- Work with resources, tools, and structured interfaces
- Understand MCP clients and servers
- Connect applications and developer tooling to AI agents
- Explore where MCP fits into modern web architectures
WebMCP
- Bring agent-accessible tools directly into web applications
- Understand how WebMCP differs from traditional MCP
- Expose application functionality to browser agents
- Use declarative WebMCP with HTML forms
- Use imperative WebMCP with JavaScript parameters
Q&A Session
Learn from the Best Speakers 2026
Meet our first speakers for 2026! More to be announced soon. Interested in joining? Apply below!
Call For Papers 2026
Companies that support the JavaScript ecosystem
Our Sponsors 2026
By becoming a sponsor of JS Poland, you gain a unique opportunity to showcase your company and products to hundreds of experienced JavaScript developers, engineers, and tech leaders from across Europe.
This kind of exposure can boost visibility, strengthen your brand within the JS community, and help you attract top engineering talent.
To learn more about our sponsorship opportunities, please fill in the form.
Special Sponsor
Special Sponsor
Special Sponsor
Special Sponsor
Special Sponsor
ngGirls Sponsor
Stay in the loop Join our mailing list
Sign up to receive updates about JS Poland, including workshops, speaker previews, ticket launches, JS Awards, JavaScript Master Podcast, Behind the Code Magazine, CFP details and other exclusive content. We won’t spam you and will only send you emails we genuinely think you’ll find interesting. You can unsubscribe at any time and you can find more information here.
- No spam
- Unsubscribe anytime
- Only the good stuff