Alun Agent Desk

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Alun Agent Desk

Inspiration

AI policy discussions often move faster than cultural institutions can verify evidence, preserve uncertainty, or include artists in decisions. Alun Agent Desk explores a simple question: what if an agent could work directly with a trusted cultural evidence surface while human judgment remained visible and accountable?

What it does

Alun Agent Desk is a shared human-agent workspace for AI-and-music developments. A person can browse source-linked records and see their evidence status, geography, implications, and uncertainty. A WebMCP-capable agent can use four structured tools to search developments, inspect a record, compare jurisdictions, and prepare an audience-specific cultural brief.

Every tool call changes the same visible interface. The person can see which evidence the agent selected, review the resulting brief, and verify the original sources. The app does not silently turn events, consultations, or platform policies into settled law.

Why WebMCP

Without WebMCP, an agent must infer meaning from page layout or conduct an unrelated web search. With WebMCP, Alun exposes a bounded, traceable evidence model directly to the agent. This makes the interaction faster and more reliable while preserving provenance and uncertainty for the human reviewer.

The key collaboration is: the agent finds and structures evidence; people retain interpretation, legitimacy, and decision authority.

How it was built

The application is a responsive React and TypeScript site hosted on ChatGPT Sites. It registers four tools through document.modelContext.registerTool: search_developments, inspect_development, compare_developments, and prepare_cultural_brief. Tool results update React state so agent actions and the human review surface remain synchronized.

The evidence set is deliberately small and source-linked. Search supports geography, status, and thematic filters, including explicit reporting when requested regional evidence is missing.

Challenges and learning

The main design challenge was preventing a polished brief from appearing more authoritative than its sources. The interface therefore keeps source URLs, evidence status, jurisdiction, and uncertainty visible. Testing also exposed a hidden audience filter that made tool results diverge from the visible list; it was removed so the agent and person now share the same result set.

What is next

The prototype can expand into the Alun Situation Room as a public cultural-technology evidence layer. Future work would add carefully reviewed records and institutional workflows without turning the tool into an automated policy authority.

Built With

• chatgpt-sites