Relay

Relay is a self-reported tool that enables Codex (an OpenAI agent framework) to interact with macOS applications by generating MCP adapters using Codex itself. The project was submitted as part of the OpenAI 2026 hackathon and is described as a proof-of-concept system built with a team of three. It claims to allow users to define application interactions via natural language, have Codex design and test an adapter, and then register it for use in agent workflows.

The description states that Relay uses Codex as its core runtime engine, leveraging Codex's ability to reason about applications and generate code for local MCP adapters. It includes a backend built with Fastify, communicates with Codex via JSON-RPC over stdio, and has a Next.js UI using Server-Sent Events for streaming turn events.

Key commercial due-diligence questions include: Is there any evidence of real-world adoption or usage beyond the hackathon? What is the actual scope of what can be integrated, and how scalable are these integrations? How does Relay handle safety and control over agent actions?

The most important open question remains: What traction, revenue, or customer data exists beyond this self-reported hackathon submission?

• The description states that Relay "turns Mac applications into tools that Codex can call through the MCP."
• It consists of two parts:
  • An adapter generation system allowing users to define intent (e.g., "Control Spotify with tools like play, pause, and skip").
  • A process where Codex designs the toolkit, writes the local MCP adapter, runs a smoke test, and registers it if successful.
• Relay includes seven reference adapters for different types of applications:
  • Minecraft
  • Clash Royale
  • Chrome
  • Messages
  • OBS
  • AppleScript
  • iMessage listener
• The system uses Codex as its runtime engine, launching codex app-server as a child process and communicating with it over newline-delimited JSON-RPC through stdio.
• It is built using Fastify backend, Next.js frontend, shadcn/ui components, and various technologies including TypeScript, React, Node.js, SQLite, and Swift.

• The description states that Codex can reason about almost any application on Mac but cannot directly interact with most of them.
• Relay is positioned as a solution to this limitation by enabling Codex to generate and install MCP adapters for applications that would otherwise require manual integration.
• The authors claim that "Codex is not just one feature inside Relay. It is the runtime behind the entire system."
• They describe how they used Codex during development itself, including one adapter completing a hackathon form submission.
• There is no evidence of prior positioning or evolution beyond this single self-reported description.

• Not evidenced.
• The description does not specify target customers or ideal customer profiles (ICP).
• No mention of who would use Relay in practice, nor whether it targets developers, end-users, enterprises, or specific verticals.

• Not evidenced.
• There is no information provided about pricing models, monetization strategies, or business models.
• The project appears to be a hackathon submission with no indication of commercial viability or revenue streams.

• Relay uses Codex as its core runtime engine.
• Communication between Relay and Codex occurs via newline-delimited JSON-RPC over stdio.
• It launches codex app-server as a child process and manages concurrent requests using a promise-chain mutex.
• Turn events are streamed to a Next.js UI using Server-Sent Events.
• Each MCP adapter runs in its own child process over stdio.
• The system supports loading new adapters without restarting the main app-server.
• Challenges mentioned include dealing with undocumented behavior, inconsistent AppleScript support, CSS selector issues, React field updates, and memory usage in headless rendering.

• Not evidenced.
• No data on user adoption, customer base, or product maturity beyond a hackathon submission.
• The project is described as a hackathon entry with no indication of post-submission traction or growth metrics.

• Not evidenced.
• No mention of existing competitors or competitive landscape.
• The description does not reference similar tools or platforms in the agent automation or MCP space.

• The entire project is self-reported and unverified; there is no independent evidence of functionality, performance, or reliability.
• The system relies heavily on Codex as its runtime engine, which may limit scalability or introduce dependency risks.
• Safety mechanisms are described as being embedded within adapters rather than prompts, suggesting potential control challenges.
• The project was submitted to a hackathon, indicating it is likely experimental and not yet production-ready.
• Lack of any revenue, customer, or traction data raises concerns about commercial viability.

1. What real-world use cases have you identified for Relay beyond the hackathon?
2. How do you plan to scale adapter generation beyond the current seven reference implementations?
3. Can you demonstrate actual functionality of the system working with non-trivial applications?
4. What are your plans for ensuring safety and control over agent actions when interacting with desktop apps?
5. Have you considered how this would integrate into existing workflows or enterprise environments?
6. Is there any plan to support other operating systems beyond macOS?
7. How do you intend to monetize or commercialize this technology?