About batuta-mcp
batuta-mcp is a stateless MCP server for splitting coding work into plans with separate file boundaries and creating a git worktree for each. It is archived and read-only; its stated purpose is to help parallel agents avoid file conflicts.
Key features
- 01Split tasks into plans with disjoint files
- 02Retry once to resolve overlapping boundaries
- 03Create an isolated git worktree per plan
- 04Preview worktree creation with dryRun
- 05Check plans for shared file boundaries
Use cases
- 01Divide a coding task among parallel agents
- 02Check planned work for overlapping files
- 03Prepare separate worktrees for each plan
- 04Preview worktree setup before writing to disk
MCP support
Exposes tools for task decomposition, boundary-overlap checks, and git worktree scaffolding to MCP clients. It uses the logged-in claude CLI for plan decomposition.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
Tools
Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.
Details
From the project README
batuta-mcp
Split a task into plans with disjoint file boundaries, then scaffold a git worktree per plan — so parallel agents never step on each other.
A stateless MCP server that turns a brain dump into conflict-free parallel work.
Retired
Claude Code now does this itself. Splitting work across isolated worktrees so parallel agents do not collide is built in, which is what this existed to provide.
The repository is archived and read-only. It still works — nothing was removed from npm, and the last published version keeps installing — but it will not be maintained, and it does not need to be. If you want the idea rather than the package, the whole design is in
docs/and the licence lets you take it.
Why
Running several coding agents in parallel is fast — until two of them edit the same file and silently clobber each other's work. The fix isn't live coordination; it's separation: give each agent a set of files that don't overlap, and the conflict can't happen by design.
batuta-mcp is the brain that does that split. You hand it a brain dump; it returns plans with disjoint file boundaries, and scaffolds an isolated git worktree for each one.
How it works
It's a stateless MCP server (stdio). No database, no daemon, no web UI — just three tools that compose:
- Decompose a brain dump into 2–5 plans whose
fileBoundariesdon't overlap (auto-corrects once if they do). - Check that the boundaries are truly disjoint.
- Scaffold a
git worktreeper plan, returning a ready-to-paste prompt for each — open them in separate terminals/tabs and let one agent work each, conflict-free.
The "muscle" (running the agents, the terminals) stays in your editor; batuta-mcp is just the planning brain.
Features
- 🧠 Disjoint decomposition — plans are generated so no file appears in two plans.
- ♻️ Auto-correction — if the model returns overlapping boundaries, it retries once to separate them.
- 🌳 Worktree scaffolding — one isolated
git worktreeper plan, with a ready-to-paste prompt. - 🔒 Safe by design — pre-flight checks (valid git repo, no overlaps, no path traversal) before touching disk;
dryRunpreviews without writing. - 🪶 Stateless — nothing persisted; uses your logged-in
claudeCLI (no API key needed).
Requirements
- Bun 1.3+
git2.x (forscaffold_worktrees)- The
claudeCLI, logged in (fordecompose_into_plans) - Claude Code or any MCP client
Installation
From npm — nothing to clone:
claude mcp add batuta -- bunx --bun @vorluno/batuta-mcp
From source — if you want to change it:
git clone https://github.com/vorluno/batuta-mcp.git
cd batuta-mcp
bun install
claude mcp add batuta -- bun run /absolute/path/to/batuta-mcp/src/index.ts
Configuration
Any MCP client works. The mcpServers entry:
{
"mcpServers": {
"batuta": {
"command": "bun",
"args": ["run", "/absolute/path/to/batuta-mcp/src/index.ts"]
}
}
}
For Claude Code, claude mcp add (above) writes this for you. For Warp or Cursor, paste the snippet into their MCP settings.
Tools
| Tool | Description |
|---|---|
decompose_into_plans | { brainDump, projectHint?, repoPath? } → { plans, overlapsResolved, attempts }. Splits the work into plans with disjoint boundaries (auto-corrects overlaps once). |
check_boundary_overlaps | { plans } → { overlaps, ok }. Pure check: do any plans share files? |
scaffold_worktrees | { repoPath, plans, dryRun? } → { results }. Pre-flight, then git worktree add per plan + a ready-to-paste prompt. |
Typical flow
decompose_into_plans→ plans with disjoint boundaries.check_boundary_overlaps→ confirmok: true(or adjust).scaffold_worktrees→ creates the worktrees; open each in its own terminal/tab and paste itssuggestedPrompt.
Development
bun test # full suite
bunx tsc --noEmit # type-check
Built test-first across 10 TDD tasks with per-task and whole-branch review.
Contributing and security
Read CONTRIBUTING.md first — its first line tells you whether your pull request will be considered. Vulnerabilities go to [email protected], never to an issue: see SECURITY.md, where the 72-hour acknowledgement is the one response time we commit to.
License
Apache-2.0 © 2026 Vorluno. See NOTICE.
Up to and including 0.1.0 this was MIT. Those releases stay MIT — a licence already granted cannot be withdrawn. From 0.2.0 on it is Apache-2.0, which grants patent rights explicitly.
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.