Agent File Workbench

Inspiration

One-off file requests are often tiny: update a name in a brief, remove comments, create a PDF packet, and send it. For an agent, that can still mean selecting libraries, installing packages, writing glue code, and checking intermediate artifacts in the user's environment.

What it does

Agent File Workbench is a local-first browser workspace for documents, PDFs, data, text, images, presentations, and media. People drop files into one visible workspace; agents use typed WebMCP tools against the same files, versions, previews, and validation results. When the deliverable is downloaded, the task workspace can be cleared in one action.

Why WebMCP

WebMCP gives the agent a stable, typed file surface instead of forcing it to infer a UI or assemble a new task-specific toolchain. The workbench exposes eight tools: workspace_file, inspect_file, extract_content, patch_document, format_document, convert_file, validate_file, and render_preview. Format-specific operations are discovered from the file at hand, so an agent can move from inspection to an auditable result through a shared browser state.

Human + agent workflow

The demo follows a last-minute client packet request. The agent lists the workspace, inspects the brief, patches a scoped section, removes comments, converts the brief to PDF, appends an approval page, validates the two-page packet, and opens the preview. The user sees each version and the final validation in the same workspace. The finished PDF is kept; task files, versions, and history are then cleared.

How we built it

The application is built with React, TypeScript, Vite, IndexedDB, and browser-native file APIs. Its WebMCP registration connects the typed tool interface to the same workspace store used by the UI, keeping human and agent operations synchronized. Document adapters provide inspect, extraction, editing, conversion, validation, and preview paths for the supported formats.

Challenges

Keeping one typed contract useful across DOCX, PDF, spreadsheets, and media required clear boundaries between inspection, mutation, preview, and validation. We also had to surface compatibility limits honestly, especially for Word-specific layout and PDF text extraction.

What we learned

The strongest WebMCP experience treats a website as a reusable task runtime: the page provides the capabilities, the agent performs structured file work, and the person keeps a visible, verifiable result.

Built With

• indexeddb
• react
• typescript
• vite
• webmcp