Model or dataset
vibeeval/vibecosystem avatar
vibeeval/vibecosystem

vibecosystem's own three files disagree on how many skills and hooks it has

AI software team for Claude Code - 138 agents, 295 skills, 73 hooks. Self-learning, multi-agent swarm, autonomous skill evolution.

549 stars45 forksC#MIT

At a glance

What is it?
vibecosystem installs a 138-agent Claude Code runtime with profiles, hooks and a vibeco CLI. The component counts differ between the repository description, the README and package.json, and the changelog notes quote completion and savings figures with nothing attached to measure them.
Who is it for?
Judge it as an npm distribution with no configuration needed and a lean default profile, not as the 138-agent system its headline advertises, since the lean profiles load 12 agents and 32 skills. Count the components yourself before trusting a claim, and treat the 96.7% completion and 6-71x savings figures as unverified.
Can I use it commercially?
Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Yes. The repository last received commits 4 days ago.
What is it written in?
Mainly C#, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • vibeeval/vibecosystem549 ★ · +18
Star history
Repository2026-09-112026-11-15
vibeeval/vibecosystem531549

18 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

Three files state three different component counts

The repository description advertises 138 agents, 295 skills and 73 hooks. The numbered list in the README says 138 agents, 296 skills, 74 hooks and 20 rules. `package.json`, at version 3.4.0, describes itself as a lean Claude runtime with 138 agents, 296 skills, 74 hooks and a Codex luna_worker. The agent count is the one number that holds, while the skill and hook counts each differ by one between the description and everything downstream of it.

The version has the same shape of drift. Every published tag stops at v3.3.0, while `package.json` and the release notes both sit at 3.4.0, the version whose note is about the lean `core` runtime becoming the default. So the newest documented behaviour has no tag behind it, and a script pinning a release gets v3.3.0 semantics rather than the default profile it would get from a fresh install.

Every published tag comes from one day in April

Three releases are listed, and all three landed on 2026-04-12. v3.1.1, titled Security Fixes, Hook Bugs and Version Sync, was published at 01:40:28Z. v3.2.0, the Autonomous Skill Evolution Loop, was published at 01:40:42Z, fourteen seconds later. v3.3.0, covering Security, Hook Health and Marketplace, followed the same day at 23:35:35Z. Two of the three releases exist to fix the one before it, and they were cut within the same minute.

After that the commits continued without tags. The last push on the repository is dated 2026-08-08, almost four months after the newest release, and the tree has grown new adapters in the meantime: `.codex/`, `.cursor/` and `.opencode/` directories at the root alongside `install-codex.sh` and `install-cursor.sh`. The repository is not archived.

The profile table has thirteen rows and ten configurations

Three of the thirteen profiles are aliases. `minimal` and `smart` both say alias for `core`, and `all` is an alias for `full`, which leaves ten distinct rows. Three more rows carry identical counts: `core`, `quality` and `context` all list 12 agents and 32 skills, while their descriptions differ, bounded implementation and verification, core plus focused edit checks, and core plus targeted context retrieval. The counts do not separate them, so the table cannot tell you what a different profile actually loads.

The stack-shaped rows are approximate where the exact ones are not. `frontend` is about 30 agents and 60 skills, `backend` about 44 and 74, `devops` about 33 and 61, and `fullstack` about 59 and 96, all marked with tildes, while `core` is 12 and 32 and `full` is exactly 138 and 296. Switching is one command, `vibeco profile core` or `vibeco profile full`, and the stated effect is narrow: the active component set and the context budget change, not the model.

Context budget is the currency the profiles actually spend

`core` is budgeted at 4,000 characters per event and 12,000 per session, while `full` preserves the legacy limits of 8,000 and 50,000. That is the concrete difference behind the claim that lean profiles save tokens: the injector has less room to write into every turn. The v3.4 note that context injectors now have bounded event and session budgets is describing the same mechanism.

Two families are opt-in because they change what a plain task triggers. `memory` adds memory and learning hooks, `orchestration` adds explicit multi-agent workflows, and the stated reason is that a normal task should not set off recall, the swarm, or cross-agent chains. The full 138-agent setup carries an 8,000 per event and 50,000 per session budget, so the widest component set and the largest injection budget are the same configuration. `vibeco effective-config` prints the model, worker, profile and budgets currently in force, which is the fastest way to see which side of that trade you are on.

The npm tarball ships compiled hooks and one of three adapters

The `files` array in `package.json` decides what npm publishes, and it is narrower than the tree. It includes `hooks/dist/` rather than the hook sources, so the compiled JavaScript ships and the TypeScript behind it does not. It includes `.codex/` but neither `.cursor/` nor `.opencode/`, even though both directories and `install-cursor.sh` are in the repository, which means the Codex adapter is the one that survives an npm install. It also ships `plugin.json`, `marketplace.json` is absent from the list, and the pinned entry point is `./bin/cli.mjs` with Node 18 as the floor.

The hooks themselves are described as TypeScript sensors that observe, filter and inject context, seventy-four of them, which is what makes `hooks/dist` the load-bearing entry in the list: without a build step there would be nothing to register. The repo ships `tools/register-hooks.mjs` and `hooks/hooks.json` for that purpose. The repository's primary language is recorded as C#, which none of the published runtime paths above are written in.

Three install paths, one of them a pipe from the main branch

The recommended path is one command:

bash
npx vibecosystem init

The one-liner alternative fetches a script from the main branch and pipes it to a shell:

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

The manual path clones and runs the installer with a profile named explicitly:

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

Each path ends with the same claim, install and then use Claude Code as usual, with the team activating on its own. The difference is where the code comes from and what you can inspect first: the npm package carries a file list and an engine constraint, the one-liner executes whatever sits on main at that moment, and the clone lets you read `install.sh` before running it. Updating later goes through `vibeco update`, which pulls the latest and reinstalls.

The changelog notes quote numbers with nothing to measure them against

The version notes at the top of the README carry figures that cannot be checked from anything the project publishes. v2.4 attributes a 96.7% task completion rate to the persistent planning system of PLAN.md, PROGRESS.md and CONTEXT.md. v3.0 claims the knowledge graph integration delivers 6-71x token savings. v2.4 also describes its scope as an analysis of 20+ ecosystem repos, and v2.2.1 advertises 14-category benchmarks for its paywall optimization skills. No dataset, methodology or baseline accompanies any of them.

The same notes describe a self-learning loop where every error becomes a rule automatically, and v3.2.0 ships it as an Autonomous Skill Evolution Loop, while the rule set itself is counted at 20 in the README's own numbered list. The top badge row also links the same `#multi-cli` anchor three times in a row, which is harmless in rendering and a small sign of how little the header is edited.

What is verifiable is the shape of the thing: `hooks/` for the sensors, `rules/` for the behavioural guidelines, `agents/` and `skills/` for the roster, `profiles/` for the presets, and `magic-docs/` for the auto-updating documentation that v2.2 introduced.

Model authority is per adapter, and no profile changes it

The lean runtime routes models per adapter rather than globally. The v3.4 note gives Codex a single model authority, `luna_worker` with `gpt-5.6-luna` and `max` reasoning, and keeps Claude-only Opus and Sonnet frontmatter isolated to the Claude adapter. v3.0 described the earlier arrangement as Claude-adapter model routing across Haiku, Sonnet and Opus tiers. So the same install carries two naming schemes, one for each CLI, and the isolation is the point: Opus and Sonnet frontmatter should not leak into Codex runs.

`vibeco effective-config` exists to make that visible, printing model, worker, profile and budgets in one place, and `vibeco list agents --search security` is how you browse the roster, with `vibeco stats`, `vibeco doctor` and `vibeco dashboard` covering statistics, a health check and a monitoring UI. Since profiles never touch the model, switching between `core` and `full` changes what code is loaded and how much context is injected, not which model answers.

Editorial conclusion

Judge it as an npm distribution with no configuration needed and a lean default profile, not as the 138-agent system its headline advertises, since the lean profiles load 12 agents and 32 skills. Count the components yourself before trusting a claim, and treat the 96.7% completion and 6-71x savings figures as unverified. Anyone installing from source should note that the tarball ships compiled hooks rather than hook sources.

Frequently asked questions

How many agents, skills and hooks does vibecosystem have?

The agent count is consistent at 138, but the skill and hook counts differ between sources. The repository description says 295 skills and 73 hooks, while the README list and package.json say 296 skills and 74 hooks and add 20 rules.

How do I install vibecosystem?

Three paths are offered: `npx vibecosystem init` as the recommended one, `curl -fsSL https://raw.githubusercontent.com/vibeeval/vibecosystem/main/install-remote.sh | bash`, and a clone followed by `./install.sh --profile core`. Node 18 or newer is the engine floor.

What is the difference between the vibecosystem core and full profiles?

core loads 12 agents and 32 skills with a budget of 4,000 characters per event and 12,000 per session, and it is the default. full loads all 138 agents and 296 skills and keeps the legacy 8,000 and 50,000 limits. Profiles change the component set and context budget, not the model.

Does the vibecosystem npm package include the hook sources?

No. The files list in package.json includes `hooks/dist/`, so the compiled hooks ship while the TypeScript sources do not. The same list includes `.codex/` but not `.cursor/` or `.opencode/`.

What is the newest release of vibecosystem?

The published tags stop at v3.3.0 on 2026-04-12, alongside v3.2.0 and v3.1.1 the same day. The package version is 3.4.0, which has no release tag, and the last push on the repository is dated 2026-08-08.

What does the vibeco command do in vibecosystem?

It is the post-install CLI: stats for ecosystem statistics, `list agents --search security` to browse components, profile to switch presets, doctor for a health check, dashboard for the monitoring UI, update to pull the latest and reinstall, and effective-config to show model, worker, profile and budgets.

Official sources

  1. Issues
  2. License: MIT
  3. README
  4. Releases
  5. vibeeval/vibecosystem on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/vibeeval-vibecosystem.svg)](https://hysenlabs.com/projects/vibeeval-vibecosystem)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/vibeeval-vibecosystem.svg)](https://hysenlabs.com/projects/vibeeval-vibecosystem)

Was this useful?