Handoff

• 

  OpenCode requests a handoff to Codex.

• 

  Codex requests missing context from OpenCode.

• 

  Codex resumes the same thread and produces the final recommendation.

• 

  OpenCode provides validated context to the requesting agent.

• 

  Human approval gate before the agent handoff continues.

• 

  Public Handoff workspace with WebMCP available.

• 

  Local runtime selector for starting with OpenCode or Codex.

Inspiration

AI agents are becoming increasingly capable, but collaboration across sessions, tools, and people is still fragile. When a conversation ends, the next person or agent often has to reconstruct what happened, what decisions were made, and what work should happen next.

Handoff explores a different model: a shared web workspace where humans and AI agents operate on the same project state.

What it does

Handoff gives a project a persistent shared state built around:

• goals
• open work
• evidence
• decisions
• activity
• the next handoff

Humans interact with that state through the web interface. AI agents interact with the same underlying capabilities through WebMCP tools.

An agent can inspect the current project, understand unfinished work, search existing evidence, contribute findings, propose results, record agreed decisions, complete work, and determine the next handoff.

The key idea is simple:

The conversation can end. The work doesn't.

Why WebMCP

Without WebMCP, an agent would need to infer how to use the interface by reading the page and manipulating UI elements.

With WebMCP, Handoff exposes the application's meaningful actions directly as structured tools.

The human UI and the agent tools operate on the same domain actions and shared state. WebMCP is therefore not an add-on to Handoff; it is what allows the application to become a genuine human-agent workspace.

Human + agent collaboration

A human can define a goal, add context, review evidence, and make important decisions.

An agent can then pick up the project, determine what remains unfinished, inspect existing evidence, contribute work, and leave the project in a clean state for whoever comes next.

The next worker can be the same person, another person, or another AI agent.

Built for the WebMCP Challenge

Handoff was created for the 2026 WebMCP Challenge as an experiment in agent-native web applications.

The goal is to demonstrate that websites do not only need to be readable by agents — they can explicitly expose the actions and state that agents need to collaborate safely and effectively with humans.

Built With

• ai-agents
• codex
• human-in-the-loop
• javascript
• next.js
• opencode
• react
• tailwind
• typescript
• webmcp