agent-ecosystem-generator

Status: alpha. Core flow smoke-tested end-to-end on a synthetic brief. Project-mode and custom-MCP paths are present but less exercised. Issues and PRs welcome.

Claude Code plugin that scaffolds a validated orchestrator + N sub-agents + M skills/MCPs ecosystem from a brief — a step up from skill-creator, which only scaffolds a single skill.

What it does

1. /generate-ecosystem spawns a requirements-interviewer sub-agent that talks to you in isolated context and produces a spec.json.
2. The orchestrator picks an orchestration pattern (orchestrator-workers, sequential pipeline, routing, parallelization-voting, evaluator-optimizer), enumerates agents/skills/MCPs, and writes plan.md.
3. You pick the output target: portable plugin or in-project .claude/ config.
4. A ecosystem-builder sub-agent materializes the files (+ scaffolds custom stdio MCPs in Python/FastMCP or TS SDK when needed).
5. A ecosystem-validator sub-agent runs semantic validation (overlap detection, pushy-ness scoring, coverage, MCP wiring) and trigger evals adapted from skill-creator. If anything fails, it loops back with fix hints up to 3 rounds.

Install

Inside a Claude Code session, register this repo as a marketplace and install the plugin:

/plugin marketplace add prgilabert/agent-ecosystem-generator
/plugin install agent-ecosystem-generator@prgilabert-plugins

Or, if you've cloned the repo locally:

/plugin marketplace add "/absolute/path/to/agent-ecosystem-generator"
/plugin install agent-ecosystem-generator@prgilabert-plugins

claude plugin install <path> (the CLI subcommand) resolves names through a registered marketplace — you can't point it at a local directory directly. The two-step flow above is the canonical path for local plugin repos until Claude Code ships first-class local installs.

Usage

/generate-ecosystem review PRs with three checkers: security, performance, tests

Or just describe what you want — the ecosystem-generator skill auto-triggers on phrases like "design a multi-agent system", "build an agent team", "scaffold an orchestrator".

Examples

See examples/ for real ecosystems produced by this plugin. Each one is an installable Claude Code plugin you can inspect or claude plugin install directly.

• examples/pr-review-trio/ — orchestrator-workers pattern: /review-pr orchestrates 3 parallel reviewers (security, performance, tests), aggregates via the pr-comment-formatter skill, references the github MCP. Synthetic reference.
• examples/space-dialogue-plugin/ — orchestrator-workers with parallel fan-out: /escribe-dialogo runs a plot writer first, then astronaut-voice + robot-voice in parallel, then a dialogue-reviewer that assembles and saves a Markdown file. Generated by a real /generate-ecosystem session (Spanish brief, 4 iterations, 0.95 validation).

Improvements over skill-creator

Feature	skill-creator	agent-ecosystem-generator
Single skill vs ecosystem	Single	Orchestrator + N agents + M skills + MCPs
Output targets	Skill dir	Portable plugin or .claude/ project config
Custom MCP scaffolding	Referenced only	FastMCP (Python) and @modelcontextprotocol/sdk (TS) templates
Validation depth	YAML fields present	Overlap detection, pushy-ness scoring, coverage, MCP wiring, trigger evals
Orchestration pattern pick	N/A	Pattern library with justification (5 canonical patterns)
Context isolation	N/A	Interviewer + builder + validator all run in isolated sub-agent ctx

See stack_explanation.md §1–6 for the background, and ROADMAP.md for what's next and where to contribute.