Replaced Dev Workflows With a GitHub PR Queue — And Now I Ship From My Phone

There is a step in most AI-assisted development workflows that nobody talks about: the handoff.

You describe what you want to an AI. It writes the code. Then you copy it, paste it into a terminal, run it, watch it work or fail, and loop back. The AI is the brain. You are the hands.

We decided to cut out the hands.

What We Built

Meridian runs on a small stack: Firebase for the consumer apps, Railway for our support platform, a Raspberry Pi 5 running our autonomous agents. Everything is wired together through GitHub.

The old workflow looked like this:

• Describe a feature or fix in a Claude.ai conversation
• Claude writes a detailed prompt
• Push it to a Notion page
• Copy the Notion page ID
• Paste it into Claude Code in the terminal
• Watch it run
• Commit and push

Every step after the description was mechanical. We were doing the work of a message queue.

So we replaced ourselves with one.

The PR Queue

The new workflow:

• Describe a feature or fix in a Claude.ai conversation
• Claude opens a GitHub PR on a feature branch with the full prompt as the PR body
• Emmett — our autonomous venture engine running on the Pi — polls open PRs every hour
• When it finds one, it runs Claude Code non-interactively against the PR body in the repo directory
• Claude Code reads the instructions, does the work, commits, and pushes
• Emmett merges the PR and logs the result
• The platform auto-deploys

The description happens on a phone. The code ships to production without a terminal ever opening.

The Infrastructure

Three components made this possible.

A GitHub MCP server on Railway. Claude.ai's custom connectors require an HTTP SSE endpoint. The official GitHub MCP server is CLI-only. We wrote a supergateway bridge that wraps it and exposes a proper SSE endpoint at a Railway URL. Claude.ai connects to it and gets full PR read/write capabilities.

Claude Code on the Pi. One install command and the Pi becomes an autonomous coding agent. It runs headless with non-interactive flags. Every PR that arrives in the queue was approved by a human upstream, so the permissions question is already answered at the moment of description.

An executor inside Emmett. A Node.js module that polls all Meridian repos for open PRs, runs Claude Code against each one with a 10-minute timeout, auto-merges on success, and closes with an error comment on failure. PRs tagged [needs-review] skip execution and surface in the morning brief instead. A cron runs it every hour. A POST endpoint lets us trigger it manually from anywhere.

What This Actually Changes

The obvious benefit is speed. A fix described on a phone ships within the hour without touching a laptop.

The less obvious benefit is discipline. When every piece of work has to be written as a PR description — clear enough for an autonomous agent to execute correctly — vague requests stop being an option. The constraint forces precision. Better prompts produce better code.

It also creates a permanent audit trail. Every change that ships has a corresponding PR with the full rationale written in plain English. The history of decisions lives in GitHub, not in memory.

The Guardrails

Autonomous execution without guardrails is how you deploy broken code at 3 AM.

A Repo Lock Protocol prevents two Claude Code sessions from running on the same repo simultaneously. PRs that touch sensitive infrastructure get tagged [needs-review] and wait for a human. The executor logs everything and surfaces failures in the morning brief.

The point is not to remove human judgment. The point is to move human judgment upstream, to the moment of description, and let everything downstream execute without friction.

The Repo

The GitHub MCP bridge is open source. The supergateway pattern works on any Railway deployment if you want to wire Claude.ai to your own repos without running anything locally.

everettsteele/github-mcp on GitHub.

Meridian is a venture studio building AI-native products. neverstill.llc