manta

⧉ Manta Clone Driven Development A new experiment in parallel AI coding. Instead of spawning cold, specialized helper agents, Manta makes Claude Code clone itself: same system prompt, your whole conversation inherited, each copy in its own isolated git worktree, all coordinating over a shared bus.

---

What it is

Subagent Driven Development spawns fresh, specialized helpers that start cold and don't know what you've been working on.

Clone Driven Development forks the agent that already understands your problem — the one you've been talking to — into N copies that work in parallel on real branches.

You're in a Claude Code session. You hit something big, repetitive, or branchy. You run one command:

manta cast forking-realities --clones 2 --task "Add rate limiting to the API"

Manta spawns 2 clones of your current agent. Each one wakes up already knowing the whole conversation, gets its own git worktree, does real work (writes code, runs tests, commits to its own branch), and coordinates with its siblings over a message bus. You review what they built and merge the best.

That's it. No role hierarchy to design, no "researcher agent + coder agent" to wire up. A clone is just you, again, somewhere else.

[!NOTE] Manta is an experiment, stated honestly: 0.1.0, early, real, not a demo. As far as we know it's the first same-system-prompt, full-transcript-inheritance cloning pattern for Claude Code. The core works and is tested (see status); the rough edges are listed, not hidden.

[!TIP] You don't have to drive the CLI. Manta looks like a power-user tool — worktrees, a message bus, file locks. You can ignore all of it. Tell your agent "orchestrate this with Manta" and it does the deciding: whether to cast, how many clones, reviewing their branches, merging the winner. You watch and approve.

Power users get the raw manta CLI. Vibe-coders get a one-line ask. Same engine underneath.

---

Why clone instead of spawn helpers?

Claude Code already has subagents (the Agent tool), and frameworks like CrewAI / LangGraph build crews of specialized roles. Manta makes a different bet:

	Subagents (Agent tool)	Role frameworks (CrewAI, LangGraph)	Manta (Clone Driven)
Context	cold — re-explain everything	per-role prompts you write	warm — inherits your live transcript
Parallelism	one-shot, returns a message	a role graph you maintain	N clones on real branches
Owns a branch + locks?	✗	varies	✓ real git worktree + bus locks
Setup	none, but starts blind	design a crew up front	none — clone what already gets it

The agent that's been in the conversation is the best worker for the task. So Manta forks that agent and lets the copies divide the work live, through coordination primitives, instead of a fixed hierarchy.

[!TIP] Task takes you 5 minutes solo? Don't cast — just do it. Manta earns its keep when the work is big, branchy, or repetitive. The built-in manta-cast-decide skill is a gut-check for exactly this.

---

How it works

The system — your main agent casts clones into isolated worktrees; they coordinate over the bus and commit to their own branches; you get a reviewed result: