vibecosystem – an AI‑software team built on Claude Code

What it is – vibecosystem is a packaged collection that turns the single‑assistant Claude Code environment into a self‑organising team of 138 specialised agents, 296 reusable “skills”, 74 TypeScript hooks and 20 behavioural rules. After a one‑line install the agents coordinate across five development phases (plan → code → review → test → learn) so you can ask Claude Code to “build a feature” and let the swarm handle design, implementation, security review, testing and post‑mortem learning without further configuration.

---

Core concepts

Concept	What it does
Agents	Markdown‑defined prompts that act as specialised roles (e.g. frontend‑dev, security‑reviewer, graphql‑expert). 138 agents are shipped.
Skills	Prompt fragments (296 total) that capture reusable knowledge such as TDD patterns, Kubernetes manifests, compliance checks, marketing playbooks, etc.
Hooks	74 TypeScript sensors that observe Claude Code tool calls, inject context and trigger agents (e.g. compiler‑in‑loop, error‑broadcast). They are loaded adaptively based on the selected profile.
Rules	20 markdown rules that shape output style, safety constraints and QA policies. New rules are generated automatically from mistakes.
Self‑learning pipeline	Errors are captured in error‑ledger.jsonl, aggregated into a skill‑matrix, and promoted to permanent .md rule files after enough occurrences across projects.
Profiles	Pre‑defined bundles that enable a subset of agents/skills to keep token usage low (e.g. core, frontend, full). Switching profiles changes the active component set and context budgets.
Dashboard & CLI	vibeco command provides stats, health checks, profile switching and a web UI that visualises agent activity and token/cost usage.

---

Quick start (npm – recommended)

npx vibecosystem init            # installs, registers hooks, creates ~/.claude layout
vibeco profile core             # load the minimal “core” bundle (12 agents, 32 skills)
vibeco dashboard                # open the monitoring UI

You can also use the one‑liner:

curl -fsSL https://raw.githubusercontent.com/vibeeval/vibecosystem/main/install-remote.sh | bash

Or clone the repo and run ./install.sh --profile core.

---

Typical workflow

You: "add a new feature"
 → Intent classifier
 → Phase 1 (Discovery): scout + architect + project‑manager
 → Phase 2 (Development): backend‑dev + frontend‑dev + devops + domain experts
 → Phase 3 (Review): code‑reviewer + security‑reviewer + qa‑engineer
 → Phase 4 (QA Loop): verifier + tdd‑guide (retries up to 3)
 → Phase 5 (Learning): self‑learner writes a rule if the task fails repeatedly
Result: feature code merged, documented and the mistake recorded for future projects.

All coordination happens through context injection – agents read and write to shared markdown files (PLAN.md, PROGRESS.md, CONTEXT.md) rather than RPC calls.

---

Profiles (token‑saving bundles)

Profile	Agents	Skills	Typical use
core (default)	12	32	Bounded implementation & verification (4 k/12 k token budget)
quality	12	32	Core + extra edit/verification checks
frontend	~30	~60	React/Next.js, CSS, accessibility
backend	~44	~74	API, DB, security
full	138	296	All capabilities, larger token budgets
all	138	296	Alias for full

Switch with vibeco profile <name>.

---

Multi‑CLI support

CLI	Install script	What you get
Claude Code	./install.sh	Full agents, skills, hooks, rules
Cursor IDE	./install‑cursor.sh /path/to/project	6 MDC rules + agents list + skills
Codex CLI (OpenAI)	./install‑codex.sh --profile core --install‑luna‑worker	Core allow‑list (32 skills) or full set, single luna_worker model authority
OpenCode	Manual copy of AGENTS.md	Skills only

---

Notable recent features (v3.4)

• Lean core runtime – default now runs the minimal runtime for speed and token efficiency.
• Claude‑adapter model routing – automatically picks Opus, Sonnet or Haiku based on the task.
• Knowledge‑graph integration – builds a code‑structure graph that saves 6‑71× tokens on context reuse.
• Dashboard v2 – real‑time token and cost tracking per agent.
• Monetization stack – agents/skills for paywall optimisation, RevenueCat patterns, Codex orchestration.
• Document generation & website cloning – agents that can generate PDFs/Docs/XLSX and clone sites pixel‑perfectly.

---

Architecture snapshot

Claude Code
 ├─ Hooks (74 TS modules) → inject context
 ├─ Agents (138 markdown prompts) ← read/write PLAN/PROGRESS/CONTEXT
 ├─ Skills (296 prompt fragments)
 ├─ Rules (20 markdown files, auto‑generated)
 └─ Memory (PostgreSQL + pgvector) – persistent per‑agent memory, smart recall

Self‑learning pipeline runs continuously: error → instinct → consolidator → rule → global promotion.

---

Data & privacy

• All artefacts live under ~/.claude/ on the local machine.
• No outbound network calls, telemetry, or cloud sync.
• Learned rules are stored locally in ~/.claude/rules/.

---

Contributing

The repo welcomes additions such as:

• New agent definitions for niche domains (e.g., Flutter, Rails).
• Skill patterns for additional frameworks or compliance regimes.
• Improved hooks for smarter context injection.
• Documentation, tutorials, and translations.

---

Bottom line – vibecosystem is a turnkey “AI software team” layer for Claude Code (and compatible CLIs) that bundles hundreds of specialised agents, skills and hooks, provides a self‑learning feedback loop, and offers a CLI/dashboard for monitoring. It aims to let developers treat Claude Code as a full‑stack development partner rather than a single, manually‑directed assistant.