A local merge queue for Claude Code agents: 90 commits a day on an 8GB MacBook Air

A developer pushing 90 commits a day from four to five parallel Claude Code agents on an 8GB MacBook Air built a tool to keep the machine from force-quitting. The tool, Claude Code Merge Queue, is a local, zero-cost merge queue that serializes agent landings so push races, redundant builds, and shared-resource test flakiness cannot happen.

The author, posting under the handle funador on Hacker News, describes the problem plainly. Running multiple agents that all try to build, test, and run dev servers on an 8GB machine is “the fast lane to a force quit and restart.” The CI bill for 90 pushes a day was also a factor. The solution is a command-line tool that runs on the developer’s own machine, costs nothing in CI minutes, and does not require a pull request.

The tool is not a theoretical architecture. It ships as an npm package. npm install --save-dev claude-code-merge-queue then npx claude-code-merge-queue init sets up numbered lanes, a FIFO queue for landings, and a pre-push hook that rejects direct pushes to the integration branch. The config file lives in claude-code-merge-queue.config.mjs. The tool writes or appends to CLAUDE.md so Claude Code lands its own work without being asked.

What the tool actually does

The merge queue provides eight commands. claude-code-merge-queue hook worktree-create plugs numbered lanes into Claude’s native worktree creation. build-lock runs a build command serialized across every lane machine-wide. land rebases and pushes a lane onto the integration branch through the FIFO queue. sync fast-forwards the main checkout and reinstalls dependencies if the lockfile changed. promote ships the integration branch to production — human-only, never automated. preview mirrors a lane’s live working tree onto the main checkout. port prints a lane’s dev-server port. prune removes already-landed sibling lane worktrees.

The emergency hatch is a single environment variable: CLAUDE_CODE_MERGE_QUEUE_EMERGENCY_PUSH=1 git push origin HEAD:main. The author calls this a convention, not a hard guarantee — it stops mistakes and stray pushes, not an adversarial agent that sets the variable itself.

The limits section is honest. No human reviews any landing. checkCommand passing is the only gate, and a real test suite and echo ok look identical to this tool. Locks are crash-safe by PID liveness, not a timeout — kill -9 anything mid-claim and the next process reclaims it. The FIFO queue lives in local temp storage, so two machines landing at once get git’s ordinary non-fast-forward rejection. A slow checkCommand caps throughput well under 20 landings per hour.

What this reveals about the agent development workflow

The tool exists because the standard development workflow does not handle four to five parallel agents. The author is not running a data center. The author is running a MacBook Air with 8GB of RAM. The fact that a single developer needs a merge queue at all — a piece of infrastructure normally associated with teams of dozens — says something about how AI coding agents change the commit cadence.

Ninety commits a day is not a human pace. It is an agent pace. The author is not typing each commit. The author is running multiple Claude Code agents that each produce code, test it, and land it. The bottleneck is no longer writing code. The bottleneck is the machine’s ability to build and test the code without crashing, and the developer’s ability to pay for CI minutes.

The tool solves the machine bottleneck locally. It does not solve the CI cost problem by being cheaper. It solves it by not using CI at all. The merge queue runs on the developer’s machine. The build and test happen there. The queue serializes them so two agents never build at the same time. The author calls it “zero-cost” because the developer already owns the machine.

The infrastructure gap

The tool also reveals a gap in the tooling ecosystem. GitHub’s Merge Queue exists, but it requires Enterprise Cloud and costs GitHub Actions minutes for every queue attempt. It also requires a pull request. The Claude Code Merge Queue requires neither. It runs on any plan, any repo. It rebases and pushes directly.

The comparison table in the README is direct. GitHub Merge Queue: private repo requires Enterprise Cloud, costs per landing, requires a pull request. Claude Code Merge Queue: any plan, any repo, zero cost, no pull request required.

This is not a feature comparison. It is a statement about what the existing infrastructure does not provide. The developer who wants to run multiple agents on a laptop cannot use GitHub’s merge queue without upgrading to Enterprise Cloud and paying for every queue attempt. The alternative is to build a local tool.

What this means for AI builders

The tool is MIT licensed. The author says to fork it, rename it, argue with the config shape. That is the point. The tool is a single developer’s solution to a problem that will become more common as more developers run multiple agents.

The limits are worth noting. No human reviews any landing. A fast checkCommand is the only throughput ceiling. The tool is not a security boundary. Every guardrail stops mistakes and convention drift, not an adversarial agent. Shell access always means git push --no-verify or editing the config on purpose.

The author is shipping a tool that works on the machine the developer already owns. That is the practical takeaway. The infrastructure for agent-scale development does not need to be expensive. It needs to be local, serialized, and crash-safe. The Claude Code Merge Queue is one developer’s answer to that need.