Forge Ai

Inspiration

Most AI web experiences stop at conversation: the agent tells you what to do, while you still have to operate the application yourself. We wanted to explore a different model—what if a website could expose its real capabilities directly to an agent?

That idea became Forge, a workspace designed from the beginning for both humans and AI agents.

What it does

Forge lets a person describe a goal in natural language, such as creating a realistic two-week product launch plan.

The agent can inspect the workspace, create and organize tasks, build milestones, modify existing work, and propose changes through WebMCP tools.

Humans stay in control: they can review agent-proposed changes, approve or reject them, and undo important actions.

Instead of an AI that only generates suggestions, Forge lets the agent actually work with the application's structured state.

How we built it

We built Forge as a web application with a visual project workspace, an integrated agent panel, structured project data, an activity log, and a human approval workflow.

The core of the project is WebMCP. Forge exposes real application capabilities through document.modelContext.registerTool(...), including project, task, planning, workspace, approval, and undo operations.

The UI and agent use the same underlying application state, so an agent action immediately becomes a visible change in the workspace.

Challenges we ran into

The biggest challenge was making WebMCP more than a technical demo. The tools had to perform real application operations and work together as a coherent workflow.

We also had to design a clear boundary between agent autonomy and human control, especially for changes that modify or remove existing work.

Accomplishments we're proud of

We built a complete human-agent workflow instead of a chatbot wrapper.

A user can give Forge a goal, the agent can use multiple WebMCP tools to work on the project, the application can show exactly what the agent did, and the human can review, approve, reject, or undo changes.

We are especially proud that WebMCP is part of the product's interaction model rather than an add-on feature.

What we learned

We learned that making a web application agent-friendly requires more than adding AI. The application's capabilities need clear structured interfaces, predictable inputs and outputs, good state management, and strong human feedback loops.

WebMCP gave us a way to think about the website itself as something an agent can understand and operate.

What's next for Forge AI

We want Forge to evolve from a project workspace into a broader agent-native environment where humans and agents can collaboratively create, organize, and execute complex work across different types of web applications.

Built With

• agents
• ai
• css
• generative
• github
• javascript
• next.js
• openai
• react
• tailwind
• typescript
• vercel
• webmcp