mpai

Product Introduction

1. Definition: mpai is an open-source, terminal-native coordination layer for AI coding assistants. It is a developer tool that enables secure, real-time multiplayer collaboration within existing, native sessions of AI agents like Codex and Claude Code.
2. Core Value Proposition: mpai exists to solve the context isolation problem in AI-assisted development. It transforms individual AI coding sessions into a shared team resource, allowing a teammate to join a live session with full context, contribute with their identity attached, and collaborate without switching tools or pasting transcripts.

Main Features

1. Native Session Integration: mpai does not create a new IDE or hosted workspace. It discovers and interfaces directly with the local session stores of supported AI agents (currently Codex and Claude Code) on the host's macOS machine. This allows it to read live transcripts and inject attributed prompts into the exact, ongoing conversation.
2. Identity-Attributed Collaboration: When a remote teammate joins a shared session and submits a prompt, their action is recorded in the native agent's transcript with their chosen name (e.g., "Alex's turn was attributed in Maya's native Codex transcript"). This preserves authorship and accountability within the collaborative workflow.
3. Owner-Controlled, Explicit Sharing: The host Mac maintains full control. Sharing is not broad or automatic. The host must explicitly create a named invite for a teammate (mpai invite --name Alex) and then share specific session IDs (mpai share SESSION_ID --with Alex). All access—list, read, prompt, presence—respects this explicit permission model.
4. Private Network Architecture: mpai leverages Tailscale to establish secure, direct connections between Macs. There is no public relay or centralized transcript cloud. All session data and communication flow over the private Tailscale network, keeping the context and collaboration within the trusted device mesh.
5. Role-Based Access & Audit Trail: Invites can be created with viewer or participant roles. The system maintains an append-only audit trail of remote prompts. This provides granular control and a verifiable history of all contributions made to a shared AI coding session.

Problems Solved

1. Pain Point: "Context Loneliness" in AI Pair Programming. A developer deep into a complex, multi-turn conversation with Claude Code cannot seamlessly hand off or collaborate within that context. Traditional methods involve screen sharing (passive), copying transcripts (lossy, out-of-sync), or re-explaining the problem (inefficient).
2. Target Audience: Technical co-founders, small engineering teams, and paired programmers who already use terminal-based AI coding tools (GitHub Copilot Codex, Claude Code) and require deep, contextual collaboration. It is ideal for remote or hybrid teams working on the same codebase.
3. Use Cases: A co-founder can directly debug a failing test within their partner's three-hour-old Claude Code session; a senior engineer can guide a junior by asking precise, in-context questions within the junior's active Codex session; two developers can collaboratively architect a solution by taking turns prompting the same AI agent with shared context.

Unique Advantages

1. Differentiation: Unlike generic screen sharing (Zoom, Discord), mpai allows active, attributed participation. Unlike shared cloud notebooks or hosted IDEs, it requires no tool change and works directly with the developer's chosen, local AI agent, preserving existing workflows and tool fidelity.
2. Key Innovation: mpai's architecture as a minimal coordination layer is its core innovation. It does not replicate session management but coordinates access to it. By sitting above the native agents and below the private network (Tailscale), it enables multiplayer functionality without owning the data pipeline or becoming a security liability. The focus on explicit session-sharing and named attribution within the native transcript is a novel approach to AI collaboration.

Frequently Asked Questions (FAQ)

1. How does mpai handle security and privacy for my AI coding sessions? mpai uses Tailscale to create a private network between trusted devices; there is no public internet exposure. The host Mac retains full control, and sharing is explicit per-session. Credentials are stored in the macOS Keychain, and the tool has no ability to execute arbitrary shell commands or delete sessions.
2. Can I use mpai with GitHub Copilot Chat or Cursor? Currently, mpai supports the terminal interfaces and local session stores of GitHub Copilot Codex and Claude Code. It does not yet integrate with IDE-embedded chat interfaces like Copilot Chat in VS Code or proprietary environments like Cursor. Support is focused on standalone, terminal-accessible agents.
3. What are the system requirements to run mpai? mpai requires macOS, an active Tailscale account and network, and either Homebrew or Node.js 20+. You must also have one of the supported AI agents (Codex or Claude Code) installed and configured for terminal use. It is currently in a public alpha state.
4. Does the remote teammate need the same AI agent (e.g., Claude Code) installed? No. The remote teammate only needs mpai installed and connected via Tailscale. The AI agent (Codex/Claude Code) runs solely on the host's Mac. The remote user interacts with the host's active session through the mpai client.
5. What happens if the host Mac goes to sleep or loses internet? The shared session becomes unavailable, as the host Mac is the source of the live AI agent session. This is a noted current boundary. Future development may address session persistence or handoff, but in the alpha, the session is tied to the host machine's state.