Enactive Brand Portal

Inspiration

Brand portals are built for human reading. When someone wants an AI assistant to apply their guidance, the usual workflow is surprisingly fragile: find the relevant page, interpret it, copy fragments into a prompt, and hope the important context survives. This especially leads to errors and failures when external suppliers, such as marketing agencies interpret the Brand Portal of their clients.

Enactive Brand Portal explores a different model. The website remains a real, human-readable brand book, while WebMCP makes the active page available to a visiting assistant as structured capability. The goal is to remove the repeated manual retransmission of brand guidance without removing human intent, factual responsibility, or judgment.

What it does

The public site presents Enactive Wonderland, an original Carroll-inspired fictional creative house. Visitors can browse its brand foundation, core voice, six character communication modes, and pen-and-ink image world like an ordinary brand portal.

On applicable pages, a compatible browser exposes Site Tools derived from that exact page.

For the canonical language journey, the person opens the Cheshire Cat page and supplies a communication goal and source facts. The assistant:

1. retrieves the active-page guidance;
2. prepares a structured Language Studio brief;
3. writes one complete draft;
4. stages the result back in the portal.

The Studio then displays the original source, proposed draft, material changes, applied rules, rationale, warnings, and ordinary human controls. The person can edit, keep, discard, copy, or reset the work. No tool approves or publishes anything.

The same pattern is implemented for visual work. The assistant can prepare and stage a generator-neutral image prompt package containing positive guidance, negative constraints, composition, alt text, and applied rules. The portal does not pretend that it generated or approved an image.

Why WebMCP

WebMCP tools change the ontological status of the brand rules from static text to executable affordances by the agent. These can be offered context dependent, so that the right tool is available on the right page of the brand portal.

Opening Cheshire Cat provides strategic reframing. Opening White Rabbit provides lucid urgency. Opening Image World provides linework, composition, colour, and provenance guidance.

WebMCP lets the assistant use that structured active-page context without DOM scraping or requiring the person to reconstruct the brand rules in every prompt. Page and session state determine which tools are available. After a copy brief is prepared, the staging tool appears as the next session affordance.

Without Site Tools, the website remains a useful brand book and manual briefing and review surface. With Site Tools, the assistant can perform the substantive transformation and return it to the same browser session for visible human review.

How we built it

Enactive Brand Portal is a static React and Vite application written in TypeScript.

Its source of truth is a four-layer JSON-LD knowledge model:

• Brand meaning;
• portal-edition content;
• generic product and WebMCP contracts;
• traceability and acceptance evidence.

The top-level page registers exactly nine browser Site Tools using:

document.modelContext.registerTool(...)

Five tools provide read-only brand and guidance access. Four prepare or stage session-local work. Local-effect tools update visible React state and sessionStorage only.

There is no remote MCP server, backend, API route, database, CMS, authentication layer, embedded model, or runtime network dependency.

Challenges I ran into

The hardest product challenge was keeping the brand and the technology separate. An earlier version technically worked but made the public experience feel like a WebMCP demonstrator rather than a useful brand portal. I rebuilt the content architecture so that Enactive Wonderland remains the subject and WebMCP appears only as a restrained application capability.

Other challenges included route-specific tool availability, dynamically adding the staging affordance after preparation, cancellation handling, output-size limits in a real host, preserving context across Studio navigation, and distinguishing native browser evidence from a local handler harness.

I also had to keep every claim honest: deterministic copy checks are partial, staged image work is a prompt package rather than an image, and the assistant proposes while the person decides.

Accomplishments

• A coherent human-facing brand portal that works without WebMCP.
• Nine page- and session-aware browser Site Tools.
• A real compatible-host copy journey covering discovery, guidance retrieval, brief preparation, dynamic staging, visible review, and reload persistence.
• A fully static architecture with immutable brand knowledge.
• Human-readable source, diff, rule, rationale, and warning inspection.
• Automated graph, schema, architecture, layer-separation, unit, accessibility, and browser checks.
• A public Vercel deployment with ten working direct routes.
• A public Apache-2.0 repository with recorded content and asset provenance.

What I learned

WebMCP is most valuable when the active page is not merely navigation but meaningful task context.

The strongest interaction was not “an agent can click the website.” It was that the website could expose purpose-built, state-aware affordances while remaining understandable and useful to a person.

We also learned that local tool execution is not equivalent to native browser proof, and that agent capability should stop before consequential human decisions such as approval or publication.

What's next

Next steps include reproducing the complete image-prompt journey in additional native hosts, testing compatibility across more WebMCP-enabled browsers, extracting the generic portal engine for other governed knowledge systems, and measuring the effect on task time, omitted rules, correction count, and reviewer confidence.

The same pattern could be applied to design systems, policy portals, learning systems, editorial standards, and other websites whose knowledge currently stops at reading and downloading.

Creator and provenance

Enactive Brand Portal is an independent experiment by Sven Spöde, a Specialist GenAI for Marketing & Communications.

Enactive Wonderland is an original fictional brand inspired by the public-domain works of Lewis Carroll. The project contains no brand assets, internal data, proprietary guidelines, or endorsement.

Research, writing, design exploration, and implementation were AI-assisted under my direction and final conceptual, editorial, architectural, licensing, and release authority.

Built With

• ajv
• axe-core
• chatgpt
• json-ld
• json-schema
• playwright
• react
• sessionstorage
• typescript
• vercel
• vite
• vitest
• webmcp