vibecosystem review: a 138-agent Claude Code team

The problem vibecosystem claims to solve

Claude Code gives you one assistant. The README states the consequence plainly: you prompt, it responds, you review, and on a complex project you end up being the planner, the reviewer, the security auditor and the tester yourself. vibecosystem is an attempt to pre-package those roles as files rather than as habits you have to build.

The target user is not a general developer. It is someone who already runs Claude Code daily and has felt the ceiling of a single conversation: no persistent memory between sessions, no automatic second opinion on a diff, no standing security pass. The repository ships 138 agent definitions, 296 skills, 74 hooks and 20 rules, and the README's own framing is that after setup you say "build a feature" and more than 20 agents coordinate across five phases. That is a claim about orchestration, and it is worth reading the rest of this review with that claim in mind.

How the agents, skills and hooks fit together

The architecture is file-based, which is the most important thing to understand before installing. Agents, skills and rules live in top-level directories (agents/, skills/, rules/) and are loaded into Claude Code rather than executed as a separate service. The hooks are TypeScript sensors; package.json ships hooks/dist/ and hooks/hooks.json, so hooks are compiled before they run.

Profiles are the control surface. Rather than loading all 138 agents and 296 skills at once, a profile selects a subset. The README's table lists core, quality, context, memory, orchestration and minimal at 12 agents and 32 skills each, with minimal as an alias for core. The larger legacy profiles (frontend, backend, fullstack, devops, full) scale up to the full set. The v3.4 release notes say the lean core runtime is now the default, which means the project has moved away from loading everything.

Model routing is split by adapter. According to the v3.4 notes, Codex uses a single model authority called luna_worker with gpt-5.6-luna and max reasoning, while Claude-only Opus/Sonnet frontmatter stays isolated to the Claude adapter. Context injectors have bounded event and session budgets, and installer ownership manifests make backups and pruning explicit. Those four sentences are the most concrete engineering detail in the release notes, and they describe a system that has been tightened rather than expanded.

Installing vibecosystem and running a first check

The README lists npm as the recommended path and requires Node 18 or later per package.json's engines field. The installer writes into your Claude Code setup, so run it from a directory where you are comfortable with file changes.

npx vibecosystem init

After that command completes, the README says the vibeco CLI becomes available. Two subcommands are worth running immediately, because they tell you what actually got installed rather than what the badge claims.

vibeco doctor
vibeco effective-config

The README describes doctor as a health check and effective-config as showing the model, worker, profile and budgets. If you want to see the component inventory before trusting it, stats and a filtered list are documented:

vibeco stats
vibeco list agents --search security

If you prefer a manual install, the README documents cloning the repository and running the shell installer with a profile flag. The core profile is the smallest documented option.

git clone https://github.com/vibeeval/vibecosystem.git
cd vibecosystem
./install.sh --profile core

The README also documents a curl one-liner against install-remote.sh. Piping a remote script into bash is a real choice with real consequences, and the README does not present it as the recommended route; npm is.

Where vibecosystem gets in your way

The first limitation is stated by the project itself. The README says the core profile covers bounded implementation and verification, and that the memory profile is the one that adds opt-in memory and learning hooks. Self-learning is in the repository description and in the solution list, but under the default profile it is not the thing you get. If you install with defaults and expect the ecosystem to accumulate rules from your mistakes, you have installed a different product than the one the description advertises.

The second issue is surface area. Even the lean core profile is 12 agents and 32 skills, plus hooks that inject context. Every one of those files consumes prompt budget and competes for the model's attention. The project has clearly recognised this, since the profile table exists specifically to save tokens and v3.4 moved the default to core, but that is mitigation, not elimination.

The third is failure mode. The README does not document rollback. There is an installer ownership manifest mentioned in the v3.4 notes that makes backups and pruning explicit, which suggests uninstall or prune behaviour exists, but the README's Quick Start does not describe how to undo an install. Before running the installer, check what the manifest backs up. If you cannot find that, treat the install as one-way.

Finally, the repository is written in C# according to its primary language field, while the hooks are TypeScript and the CLI is an .mjs file. The README does not explain what the C# portion covers. That is an unresolved question, not a defect, but it means you should not assume the whole stack is one language.

vibecosystem compared with writing your own Claude Code setup

The real alternative is not another product. It is hand-rolling the same thing with Claude Code's own extension points: a handful of agent markdown files, a couple of hooks, and your own rules. That approach gives you exactly the roles you need and nothing else, and it is fully legible because you wrote it.

The difference in approach is scale and maintenance. A hand-rolled setup of five agents is easy to reason about and easy to debug when an agent behaves oddly. vibecosystem ships 138 agent definitions and 296 skills, which means the value is in the breadth you did not have to author, and the cost is in the breadth you did not author. When something misfires, the search space is larger. The project's own answer to this is the profile system and the vibeco CLI, which is a genuine architectural response rather than a marketing one: profiles bound the blast radius, and doctor and effective-config give you a way to inspect what is loaded.

A second alternative is to use vibecosystem only for the parts you cannot easily write, such as the security and review agents, and keep your own planning workflow. The profile table supports that, since core, quality and context are all 12 agents and 32 skills, letting you pick a narrow slice without adopting the full ecosystem.

Licence, maintenance and upgrade cost

vibecosystem is MIT licensed, which permits commercial use, modification and redistribution with the licence and copyright notice preserved. That is permissive and carries no copyleft obligation on your own code. This is a description of the licence text, not legal advice; if you are redistributing the ecosystem inside a product, read LICENSE and consult your own counsel.

On maintenance, the last push to the repository was on 2026-08-08, roughly six weeks before this review. The most recent release listed is v3.3.0 from 2026-04-12, titled "Security, Hook Health, Marketplace". The README and package.json describe v3.4, including the lean core runtime as the default, but no v3.4 release entry appears in the release list, so the README is ahead of the tagged releases. That gap is the thing to check before you pin a version.

Upgrade cost is shaped by the profile system. Because profiles are the unit of loading, moving between them changes what is in context, and the v3.4 notes mention installer ownership manifests for backups and pruning. The README documents vibeco update as pulling the latest and reinstalling. Between v3.0's knowledge graph integration and v3.4's bounded context injectors, the direction of travel is toward fewer things loaded by default, which means an upgrade may silently change which agents are active. Run vibeco effective-config after any update to confirm the profile and budgets you expected.

Editorial conclusion

Adopt vibecosystem if you already live inside Claude Code and want role separation without writing your own agent definitions. Skip it if you need a single stable toolchain or cannot tolerate a large prompt surface. Before installing, run npx vibecosystem init, then vibeco doctor and vibeco effective-config, and check the installer ownership manifest before letting it prune anything.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. Releases
5. vibeeval/vibecosystem on GitHub