Claude Code Plugin

Parallel Agent Swarms for Code

Deploy fleets of specialized AI agents that scout, plan, implement, and review your code—all running in parallel.

$ /plugin install cas click to copy
Works with Claude Code — zero config required
Explore

Ten skills.
Infinite leverage.

Each skill orchestrates a different swarm topology—from intelligent routing to iterative autonomous loops and adversarial verification.

zk
Intelligent Router

Analyzes your task and automatically picks the best skill—Siege, Legion, Hydra, PCC-Opus, or PCC—so you don't have to.

auto-routing
pcc
Parallel Claude Coordinator

Sonnet scouts explore the codebase in parallel, then Opus implementers execute the plan with precision.

scout + implement
pcc-opus
PCC Opus Edition

Maximum exploration quality. Opus-powered scouts for when the codebase is complex and nuance matters.

max quality
siege
External Orchestrator

Spawns fresh claude -p sessions per iteration with two-skeptic adversarial verification. Arithmetic-only exit decisions.

beta — max rigor
legion
Iterative Swarm Loop

Submit a holistic project. Legion keeps deploying swarms autonomously until everything is built or progress stalls.

beta — autonomous loop
hydra
Multi-Task Swarm

Submit N tasks at once. Agents collaborate via mailbox messaging within waves. Two-skeptic adversarial global verification.

beta — n-task parallel
review
7-Agent Code Review

Seven specialized review agents analyze bugs, style, silent failures, comments, types, and test coverage simultaneously.

7 agents
systemcc
Auto-Routing Workflow

Lyra AI optimization with triple review gates. The full pipeline—routing, implementation, and validation.

full pipeline

Orchestrate, don't operate.

You give the task. CAS splits it into parallel workstreams, assigns specialized agents, and synthesizes their output.

1
You Command

Run a skill like /pcc with your task description.

2
Scouts Explore

Parallel scout agents map the codebase, finding patterns and relevant files.

3
Agents Execute

Implementer agents work concurrently, each owning their file set exclusively.

4
Output Synthesized

Results are deduplicated, cross-referenced, and presented as one cohesive deliverable.

Hydra swarm topology.

The multi-headed orchestrator uses Claude Teams—a shared task list, message passing, and wave-based execution to run N tasks with zero file conflicts.

you/hydra task1; task2; task3
TeamCreate
Hydra OrchestratorOpus — brain, not hands
spawns in parallel
parallel — all at once
scoutarchitecture
scoutfeature
scoutdependency
scouttest
scoutintegration
SendMessage → findings
analyst-synthesisreads all scouts → plans + conflict DAG
wave schedule computed
wave 1
parallel — mailbox collaboration + exclusive file ownership
impl-task1stream-a
impl-task1stream-b
impl-task2stream-a
impl-task2stream-b
verify → pass → next wave
wave 2
blocked until wave 1 verified
impl-task3stream-a
impl-task3stream-b
all waves complete → two-skeptic debate
skeptic-a-globalindependent evaluation
skeptic-b-globaladversarial debate
simplify-module-across-task cleanup
simplify-module-bcross-task cleanup
TeamDelete → done
Final ReportN tasks • W waves • ~27 teammates • collaboration metrics
Explore agents
Analyst
Implementers
Two-Skeptic Verifiers
Simplifiers
SendMessage (inter-agent)

Ready to deploy the swarm?

Install in one command. No configuration needed.

View on GitHub