A focused runtime for Claude Code. A bounded worker for Codex.
Catalog (in repo): 
Default core:
Türkçe · Español · Français · Deutsch · Português · Italiano · Nederlands · 中文 · 日本語 · 한국어 · العربية · हिन्दी · Русский
vibecosystem packages agents, skills and hooks for software work. The default core profile installs a selected set for investigation, implementation, review and verification. Memory, learning and orchestration have explicit profiles; full installs all 138 agents, 296 skills and 20 rules plus 53 of the 74 hooks.
3.5.1 is the current npm release. See CHANGELOG.md and the upgrade guide.
Use a source checkout to try these changes. When reviewing a PR, check out its head branch before installing; cloning uses main by default:
git clone https://github.com/vibeeval/vibecosystem.git
cd vibecosystem
./install.sh --profile core --dry-run
./install.sh --profile coreThe installer uses the prebuilt hooks in hooks/dist/; a normal installation does not compile TypeScript or install the hook development dependencies. Node.js 18+ and Bash are required. Development checks use Node.js 22.
For the published npm version (currently 3.5.0):
npx vibecosystem initThe remote installer fetches main (this source tree). Review the script before running it:
curl -fsSL https://raw.githubusercontent.com/vibeeval/vibecosystem/main/install-remote.sh -o install-remote.sh
bash install-remote.sh --profile coreRestart Claude Code after changing its hook registrations. Inspect the installed state with:
vibeco effective-config
vibeco doctor
vibeco statsvibeco effective-config shows which profile is active and how many hooks are registered and compiled.
The source installer creates ~/.local/bin/vibeco unless an unmanaged launcher already exists there. Add ~/.local/bin to PATH if your shell cannot find it. You can also run the source CLI with node tools/vibeco/vibeco.mjs.
Profiles select installed components and registered hook commands. Catalog size does not describe how much prompt content is loaded into each request.
| Profile | Purpose |
|---|---|
core |
Selected implementation, investigation and verification tools; default |
dashboard |
Core plus opt-in local activity monitoring |
quality |
Core plus focused edit, compiler and security checks |
context |
Core plus targeted code context retrieval |
memory |
Core plus persistent recall and learning |
orchestration |
Core plus explicit multi-agent workflows |
full |
All 138 agents, 296 skills and 20 rules; 53 of the 74 hooks |
minimal and smart alias core; all aliases full. Legacy frontend, backend, fullstack and devops profiles retain domain allowlists with core hooks. Read the exact selection in profiles/runtime-manifest.json, or use vibeco effective-config after installation.
vibeco profile quality # core plus focused edit, compiler and security checks
vibeco profile core # default selected set
vibeco profile memory # core plus persistent recall and learning
vibeco profile full # all agents, skills and rules; 53 of 74 hooksProfile changes reconcile installer-owned files and hook registrations. Modified or unmanaged files require review; the installer preserves ownership boundaries and creates backups before replacements. See UPGRADING.md.
The default profile avoids installing the complete agent and skill catalog. Context-producing hooks use character limits: 4,000 per event and 12,000 per session for lean profiles; 8,000/50,000 for full. These are limits on participating hook context, not limits on total model input, output, reasoning, tool results or billing.
Token cost depends on the selected model, task, retrieved content and retries. Character counts and estimated_tokens are estimates; provider-reported usage is recorded separately when present. This project does not promise a fixed percentage or multiplier of token savings. Reproducible measurements belong with their fixtures and commands in the verification record.
The core skill entrypoint payload is smaller after focused rewrites: see the reproducible byte comparison. Skills load on demand, so this is an available instruction size measurement, not a whole-session token claim.
For routine work, stay on core, inspect relevant files, use one bounded implementation path and run the checks required for the change. Select context, memory or orchestration when the task benefits from them. Installing full does not automatically launch an agent swarm.
flowchart LR
P[Selected profile] --> I[Installer]
I --> C[Selected agents and skills]
I --> H[Registered Claude hooks]
E[Claude lifecycle event] --> H
H --> B[Bounded optional context]
B --> W[Claude session]
C --> W
W --> V[Implementation and verification]
Hooks run in response to host lifecycle events. They can inspect tool input, return a permission decision or provide context. They do not guarantee that a suggested agent or workflow will execute. Agent and skill files are instructions consumed by the host model.
The optional memory pipeline captures and consolidates patterns under the user's Claude directory. Its thresholds are heuristics, not proof that a learned rule is correct. Review learned material before relying on it in sensitive work.
| Path | Catalog (in repo) | Default core |
|---|---|---|
agents/ |
138 Markdown agent definitions | 12 |
skills/ |
296 top-level skill directories; some contain nested skills | 32 |
hooks/src/ |
74 TypeScript entry points, including utilities and CLIs | 6 |
rules/ |
20 rule documents; the lean profiles do not install the full rule set | - |
The full profile installs 138 agents, 296 skills, 20 rules and 53 of the 74 hooks.
Source entry points, unique hook names and registered commands are different counts. hooks/hooks.json is the registration source. Some files are command-line utilities or are not registered by default. Use vibeco list agents --search security to browse the installed catalog.
| Host | Setup | Scope |
|---|---|---|
| Claude Code | ./install.sh --profile core |
Selected agents, skills, hooks and rules |
| Codex | ./install-codex.sh --profile core --install-luna-worker |
Selected skills and an optional bounded luna_worker |
| Cursor | ./install-cursor.sh /path/to/project |
Project instructions, MDC rules and skills |
| Other compatible hosts | Manual | Skill content; check the host's format and discovery rules |
The Codex worker contract is .codex/agents/luna-worker.toml: gpt-5.6-luna with max reasoning. The project config stays model-neutral. Claude-only opus/sonnet metadata is not applied by the Codex installer. Model availability and custom-agent support depend on the host and account. See docs/codex-setup.md.
Core installation does not provision external services. Some optional hooks and skills need tools that are not bundled:
- Code context and Python diagnostics use a compatible TLDR daemon; unavailable diagnostics can return no findings.
- Memory graph features need a Node-compatible
better-sqlite3installed in~/.claude/hooks/; a missing or incompatible module is reported and that graph operation skips. Other recall workflows reference Python/uvscripts and may need a configured database. - The
dashboardprofile adds one emitter to core (7 unique hook entry points across 8 registrations). It records agent starts on PreToolUse and structural tool completion events on PostToolUse;fullincludes it too. The dashboard uses the localwspackage. Install it withnpm ci --ignore-scripts --prefix ~/.claude/tools/dashboard, as described in its setup guide. - Compiler assistance may contact a configured LM Studio endpoint. Braintrust extraction requires
VIBECO_BRAINTRUST_ANALYSIS=1, a key and its script. GitHub-mutating hooks requireVIBECO_GITHUB_AUTOMATION=1. MCP, search and provider skills may contact their configured services.
Local state can contain code excerpts, prompts, tool output, errors and learned rules under ~/.claude/. The core product is local; integrations may send data to the services you configure. Review the enabled profile, credentials and endpoint settings. The credential hook is a best-effort guard and does not replace host permissions or an operating-system sandbox. See SECURITY.md.
cd hooks
npm ci
npm run check
npm run build
npm testKeep hooks/dist/ synchronized with source. Installer tests use temporary home directories; do not use a personal installation as a fixture. See CONTRIBUTING.md, ARCHITECTURE.md and docs/verification.md.
The current eight-second GIF is a locally rendered product diagram, not a screen recording or benchmark. Its editable source freezes data from the runtime manifest. The static diagram describes profile selection and runtime boundaries. Earlier terminal GIFs remain in assets/ as historical material; they depict legacy workflows and are not evidence that every feature runs in the default profile. See assets/README.md for sources and reproduction requirements.
vibecosystem, Claude Code için agent, skill ve hook'ları bir araya getirir. Varsayılan core profili, araştırma, geliştirme, inceleme ve doğrulama için seçilmiş bileşenleri kurar. Bellek, öğrenme ve çoklu agent akışları ayrıca seçilir; katalogdaki her şey her oturumda yüklenmez.
./install.sh --profile core --dry-run
./install.sh --profile core
vibeco effective-configvibeco effective-config aktif profili ve kayıtlı hook sayısını gösterir.
Token verimliliği için önce göreve uygun küçük profili seç. Bağlam sınırları modelin toplam token tüketimini sınırlamaz; ücret ve kullanım görevine, modele ve tekrar sayısına bağlıdır. Ölçülmemiş tasarruf oranları vaat etmiyoruz. full profili 138 agent, 296 skill ve 20 rule ile 74 hook'tan 53'ünü kurar; otomatik olarak bir agent ordusu başlatmaz.
3.5.1 şu anda npm üzerindeki güncel sürümdür. Yayınlanmış sürüm ve kurulum için İngilizce bölümü, değişiklikler için CHANGELOG.md dosyasını takip edebilirsin.
Built by @vibeeval. MIT licensed. See LICENSE.
The catalog incorporates ideas credited in the individual skills and agent definitions, including work from Shannon, UI UX Pro Max, Game Studios, Skill Gateway, Pyxel, copilot-orchestra and Trail of Bits Skills. External tools and providers retain their own licenses and terms.