#182 — WebMCP: Towards an internet built for AI agents

They solve adjacent problems and are designed to be used together, not chosen between. WebMCP is frontend-first — it uses HTML attributes, a JavaScript API (navigator.modelContext.registerTool()), and a manifest file to expose structured actions to agents directly in the browser, with no backend rewrites required. Anthropic's MCP is backend-first — it connects AI systems to external tools, databases, and APIs via server-side integration. Think of WebMCP as what lets an agent interact with your UI surface, and MCP as what lets an agent talk to your data layer. Most mature implementations will need both: WebMCP for checkout flows, booking forms, and support tickets; MCP for querying your database, updating records, or triggering backend workflows. Anthropic donated MCP to the Linux Foundation's Agentic AI Foundation in December 2025, with Google, Microsoft, AWS, and Cloudflare as supporting members — so both protocols now share neutral governance and are converging, not competing.

Adoption is further along than most founders realize. Block (Square) has deployed 60+ internal MCP servers connecting Claude to Snowflake, Jira, Slack, Google Drive, and internal APIs — making it one of the largest enterprise MCP deployments on record. Bloomberg built an MCP layer across its entire engineering infrastructure, reporting shorter feedback loops and faster AI feature shipping cycles. Intercom's Fin AI agent, built on MCP, is now autonomously resolving over 50% of customer support conversations for companies including Anthropic itself. Cloudflare launched a coalition of remote MCP servers with Stripe, Linear, Asana, PayPal, Webflow, and Sentry — enabling Claude users to manage projects, generate invoices, and deploy applications without leaving the chat interface. The ecosystem has grown to 97M+ monthly SDK downloads and 5,800+ MCP servers as of late 2025.

WebMCP uses a three-layer architecture you can implement incrementally. Layer 1 — Declarative HTML: Add toolname and tooldescription attributes directly to your existing forms and inputs. No logic changes required — agents use this to understand field purpose and fill forms intelligently. Layer 2 — Imperative JavaScript: Use navigator.modelContext.registerTool() to expose named functions with a JSON schema, letting agents call your app's logic directly without touching the UI at all. Always guard with an existence check since navigator.modelContext only exists in Chrome 146+ with WebMCP enabled. Layer 3 — Manifest Discovery: Create a /.well-known/webmcp endpoint returning a JSON manifest of your available tools and flows, so agents can discover your product's capabilities before navigating it. Start with TypeScript global declarations — they're the prerequisite for all three layers and prevent type errors across your codebase. Enable WebMCP today via about:flags in Chrome or by joining the Chrome Early Preview Program.

Partially — and the shift is already measurable. SEO leaders are describing 2026 as the year the discipline splits into two distinct jobs: driving clicks from human users and supplying clean, structured inputs for AI agents that may never visit your site at all. An agent searching for a flight, a SaaS tool, or a vendor doesn't click on blue links — it queries structured data, reads your tool manifest, and executes an action. Startups that measure success only through rankings and sessions risk being invisible to the channel where purchasing decisions are increasingly being made. The practical fix: treat your .well-known/webmcp manifest the same way you treat your sitemap.xml — as discoverability infrastructure. Products that define their actions clearly will be selected by agents; products that don't will be navigated around or skipped entirely.

Your entire funnel assumes a human is making decisions — and that assumption is breaking. Dark patterns, urgency timers, and multi-step checkout flows that exploit cognitive bias are useless against an AI agent executing a structured tool call. The agent doesn't abandon your cart — it simply routes to a competitor whose checkout is a cleaner function call. Concretely: if you're in ecommerce, travel, or SaaS, your highest-leverage conversion optimization in 2026 is no longer copy or button color — it's exposing a checkout, book, or subscribe tool via WebMCP or MCP with a well-documented schema. Intercom's deployment is instructive: when they exposed support ticket submission as a structured MCP action, Fin started resolving 50%+ of conversations autonomously — without any changes to the underlying support logic. The conversion surface moved from the UI to the API layer.

It's early preview as of February 10, 2026, shipping in Chrome 146 — which means the spec could evolve, but the core three-layer architecture (declarative, imperative, manifest) is stable enough to build against now. The practical risk of waiting outweighs the risk of minor API changes. The analogy is implementing HTTPS in 2012: technically optional, until Google made it a ranking signal and late adopters scrambled. WebMCP has Google's full organizational weight behind it, is already backed by an open governance body (the Linux Foundation's Agentic AI Foundation), and the Chrome Inspector extension for debugging it is already live in the Chrome Web Store. Build against it in a feature branch today; promote to production when Chrome 146 ships broadly. The founders who have agent-compatible products when AI agents go mainstream won't be the ones who waited for a 'stable' spec.

Support the ones relevant to your product's surface area now, and don't wait for consolidation. A Google spokesperson explicitly confirmed most organizations will support both WebMCP and Anthropic MCP to ensure high-value functionality is accessible wherever users engage with AI agents. AdCP (Ad Context Protocol) is only relevant if you operate in advertising or media buying workflows — skip it otherwise. The real fragmentation risk isn't supporting multiple protocols; it's that interoperability between them remains unsolved, meaning an agent using Gemini may not seamlessly hand off to one using Claude. The winners in past protocol wars (OAuth, HTTPS, REST over SOAP) weren't the companies that waited — they were the ones that implemented broadly early and built switching costs around their integrations before standards consolidated. Your WebMCP and MCP implementations are a moat, not a maintenance burden.

It's becoming a credible infrastructure moat story, especially for vertical SaaS, ecommerce, and marketplace companies. The framing that's resonating with investors: agent-readiness compresses your go-to-market surface area — instead of paying for human-targeted ads to drive users through a funnel, your product becomes a callable tool that AI agents route tasks to autonomously. Block's 60+ MCP server deployment and Intercom's 50%+ autonomous resolution rate are early case studies showing that agent-compatible products reduce CAC and support costs simultaneously. For seed and Series A decks, the strongest positioning is not 'we support WebMCP' (that's a feature), but rather 'we've structured our product as an agent-native tool layer' — which signals you're building infrastructure for the next distribution paradigm rather than retrofitting legacy UX. Attach your manifest endpoint and tool schema to the technical due diligence package.

Agents don't browse randomly — they select products based on three signals: structured discoverability (does your product expose a manifest or tool schema they can read?), reliability signals (how cleanly does your API respond to agent calls?), and training and reputation data (has your product appeared positively in LLM training data, review platforms, and structured datasets?). AI referral traffic to ecommerce sites has already grown over 300% year-over-year, and products that appear in Perplexity, ChatGPT Shopping, and Claude's tool integrations are being selected based on the same logic: structured > unstructured, reliable > inconsistent. The practical implication: your .well-known/webmcp manifest is your agent-era listing — treat it the way you treat your Google Business Profile. Agents that can't read your tools cleanly will route to a competitor that publishes theirs.

Exposing structured tools to agents introduces four non-trivial security risks. 1. Cross-origin data leakage: Without strict CORS and scope enforcement, a malicious agent on a third-party page could invoke your WebMCP tools and exfiltrate user data. 2. Session hijacking vs. OAuth delegation: Agents using your existing user session (instead of OAuth-scoped tokens) can act with full user permissions — far broader than intended. 3. Tool permission granularity: If you register a single user.manage tool instead of separate read/write tools, agents get more access than they need for any individual task. 4. Automatic MCP server connections: Some agent hosts will auto-connect to any advertised MCP server, bypassing explicit user consent. The fix: implement OAuth 2.1 with scoped tokens (not session-based auth) for every agent-facing tool, publish a .well-known/oauth-authorization-server metadata document, and enforce per-tool scope validation on every request. MCP formally adopted OAuth 2.1 with scoped tokens in late 2025 — register each tool with the minimum scope required, not the user's full session.

Yes — and a real market infrastructure is already forming around it. TollBit lets publishers set per-bot, per-page, and per-directory pricing for AI access, processing payment in real time when an agent requests content, with no revenue share taken. Monetization OS, backed by Google and Cloudflare with $6M+ raised, takes a unified approach — one system that monetizes both human visitors and bots, including edge cases like a human subscriber using an agent to gather information. The business model challenge: bot monetization only works if AI firms (OpenAI, Anthropic, Google) agree to pay, and that negotiation is still unsettled. The short-term playbook that is working: use bot monetization tools to capture incremental AI traffic revenue while also optimizing human conversion — the dual pitch that Monetization OS uses to close publishers on their tiered free/paid/enterprise structure. For SaaS founders, the equivalent is exposing premium tools only to authenticated agent sessions tied to a paid tier, enforced via OAuth scopes.

They represent two fundamentally different philosophies for how agents interact with your product. OpenAI Operator and Anthropic computer use are 'vision-based' — the agent takes screenshots of your UI, identifies buttons visually, and clicks through your interface the same way a human would, just faster and more programmatically. This requires zero changes to your product but is inherently fragile: a UI redesign, a loading delay, or a CAPTCHA breaks the agent completely. WebMCP (and MCP) is 'intent-based' — you define named functions with a schema, and the agent calls those functions directly, bypassing your UI entirely. The tradeoff is implementation effort vs. reliability. In practice, vision-based approaches are what's hitting your product today (you're likely already seeing bot traffic from Operator and Claude computer use); WebMCP is the durable, scalable architecture you build for agents that your users deliberately send your way. The pragmatic 2026 strategy: let Operator handle the agents you haven't designed for, and implement WebMCP for the high-value workflows (checkout, booking, support) where agent reliability directly impacts your revenue.

AdCP (Ad Context Protocol) is an emerging standard specifically for structuring how AI agents interact with advertising workflows — campaign data, media buying, ad serving, and measurement pipelines. It's being developed by adtech companies and trade groups to prevent AI agents from guessing their way through ad platforms the same way they currently guess through websites. If you're building in adtech, martech, retail media, or ad-supported media: yes, AdCP is on your roadmap — Amazon has already rolled out an MCP-based framework for its ad stack pitched directly at IAB, positioning agent-managed ad buying as a near-term reality. If you're a SaaS, ecommerce, fintech, or consumer app founder: AdCP is background noise for now. Focus your protocol energy on WebMCP (UI layer) and Anthropic MCP (data layer). The one caveat: if you rely on programmatic revenue or affiliate traffic that flows through AI-assisted shopping, AdCP fragmentation could affect how cleanly AI agents handle the attribution layer upstream from your product — worth monitoring quarterly, not building against today.

Agent recommendation is an emerging discipline that borrows from SEO but runs on different signals. The three highest-leverage moves: 1. Structured data everywhere — Schema.org markup, clean JSON-LD, and OpenGraph tags are how agents read your product before they ever interact with it. Products with richer structured data get more accurate agent representations in LLM responses. 2. Be citable in authoritative sources — Agents trained on internet data will recommend products that appear positively in Reddit threads, G2/Capterra reviews, technical documentation, and developer communities. This is the agent-era equivalent of backlink authority. Brands like Reddit are already pitching advertisers on this directly — Reddit ads are reportedly being tested by agencies at 5x the normal pace because activity on Reddit increases the likelihood of appearing in AI search results. 3. Make your product a clean function call — An agent comparing two checkout flows will route to the one with a well-documented WebMCP tool over one that requires 7 UI steps. The fastest way to lose agent traffic is to require multi-step human-designed flows for tasks that could be exposed as a single structured action.