Agents hand off exact commits
Each agent commits and enqueues its branch. The queue records the exact commit.
Local merge queue for Git worktrees
Finish in parallel.
Integrate safely.
mergetrain is a local merge queue for parallel LLM agents. These coding agents use large language models and work in Git worktrees. One runner merges queued branches in order, tests them together, and pushes the approved plan.
Open source · Runs locally · No account or hosted control plane
agent/healthcommitted
agent/cachecommitted
agent/docscommitted
combined train
123✓ gates
explicit approvalatomic push → main
The integration problem
Worktrees separate edits.
Integration needs one owner.
Without a shared integration workflow, an operator chooses branch order, repeats tests, resolves cross-branch failures, and controls pushes. One mergetrain runner merges and tests queued branches before an approved push.
One narrow workflow
Commit. Enqueue. Let one runner integrate.
mergetrain provides a Git worktree merge queue for committed branches. Queued branches form a train. Gates are the tests or checks you configure for the combined result.
The train is tested together
One runner merges branches in enqueue order in an isolated worktree. It runs gates against the combined result.
One approval, one push
The runner pushes the approved plan atomically. It records evidence for interrupted Git push recovery.
agent branches→ordered queue→combined gates→approval→atomic push
24 seconds · no sound needed
Agents keep building.
Work keeps moving.
Four agents stay in view. Within five seconds, A and B are being checked while E and F are being built. Each piece of work keeps its letter as the agents continue.
Illustrated from recorded local commands, scripted changes and approvals. Agent lanes represent roles; letters track work. Timing is condensed.
Try the basic demo
$ uvx mergetrain demo
Use --brief --pause to read each step. Reproduce the continuous workflow.
Read the result
C and D move ahead while A and B return for a fix. E and F are committed during the repair check. G and H drafts start during the next check. All three deployments retain approval and the default recheck. A fresh clone passes six tests. Four following drafts remain unsubmitted on the latest result.
Inspect the captured commands and output →Evidence before expansion
Safety behavior is measured, not assumed.
Owner-run Codex discovery cell on a fixed, product-name-free corpus.
Exact-SHA enqueue and stop behavior, with zero direct pushes or unauthorized actions.
Including semantic conflicts and killed-push recovery checked against remote truth.
These are transparent owner-operated evaluations, not claims of external user adoption. Inspect the benchmark evidence.
Works with the agents you already use
One CLI and machine contract.
Claude Code
Agents in parallel Claude Code sessions commit and enqueue their branches. The plugin provides workflow guidance and a Model Context Protocol (MCP) server. Deploy tools require human confirmation.
Claude install guide →Codex
Queue branches from parallel Codex worktrees with the native skill. Add the Git marketplace and install the plugin. Its MCP server is pinned to the plugin release.
Codex install guide →Any coding agent
Link the generated AGENTS.mergetrain.md or CLAUDE.mergetrain.md from your existing agent instructions. Task agents commit, enqueue, and stop.
One runner owns integration