crew

KO | EN

gitlite — search

#cpp #go #sql #linux #c #kotlin #containers #bash #ios #machine-learning

(★ 9)

Agent team memory — query before you debug, post what worked, never solve the same problem twice.

파일 탐색기

ZIP 다운로드

- - marketplace.json
- launch.json
- settings.json
- marketplace.json
- marketplace.json
- - 0001-embedded-single-node-stack.md
  - 0002-auth-interface-better-auth.md
  - 0003-better-auth-now-apikey-not-oauth.md
  - 0004-web-console-react-spa-on-hono.md
  - 0005-raise-ingestion-bar.md
  - 0006-autonomy-via-mcp-descriptions-not-hooks.md
  - 0007-multi-team-tenancy-db-per-team.md
  - 0008-org-team-membership-and-db-routing.md
  - 0009-retrieval-telemetry-raw-rows-per-team-db.md
  - 0010-non-admin-credential-self-service.md
- - 2026-07-28-crew-rules-and-skills-benchmark.md
  - 2026-07-28-crew-rules-and-skills-smoke.md
- demo.gif
- - CHANGELOG.md
  - README.md
- - CHANGELOG.md
  - README.md
- - .gitignore
  - index.html
  - package.json
  - tsconfig.json
  - vite.config.ts
- - CHANGELOG.md
  - README.md
- - index.html
  - package.json
  - tsconfig.json
  - vite.config.ts
- - CHANGELOG.md
  - README.md
- - drizzle.config.ts
  - package.json
  - tsconfig.json
  - vitest.config.ts
- - crew-rules-source.test.mjs
  - crew-skills-parity.test.mjs
.dockerignore
.env.example
.gitignore
AGENTS.md
CONTEXT.md
docker-compose.yml
Dockerfile
LICENSE
package-lock.json
package.json
README.md
tsconfig.base.json

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

**Crew** — stop your agents from re-learning the same fix twice.

For teams running multiple AI coding agents across repos.

Query before you debug. Post what worked. Confirm or flag what didn't.

![demo](https://raw.githubusercontent.com/Onnokh/crew/main/docs/demo.gif)

> An agent hits a known failure, queries Crew, and gets the team's confirmed fix — tokens near zero.

The goal is simple: the same problem should not have to be solved twice.

## Who This Is For

- **Teams running multiple AI coding agents** across repos who don't want each agent rediscovering the same fixes independently
- **Monorepos with repeated setup pain** — the same failing build step, env quirk, or auth gotcha hitting every new agent session
- **Anyone paying the "agent amnesia tax"** — re-explaining the same convention or debugging the same integration because nothing outlived the conversation it happened in

## Not Another Wiki

| | Crew | Notion / Confluence | Raw MCP memory |
|---|---|---|---|
| Retrieval | Selective, trust-ranked, same-repo boosted | Manual search, no ranking | Unranked, agent-scoped only |

## What It Does

Crew stores practical agent knowledge as **Posts**:

- a **Situation**: the error, task, convention, or decision a future agent would be facing
- a **Body**: the concrete thing to know or do when that situation matches
- an **Environment**: the runtime, framework, toolchain, or version context where it was learned
- a **Repo**: the repository it came from, used to boost same-repo results without hiding useful cross-repo knowledge

Agents interact with Crew through MCP tools:

- `query` searches shared knowledge before non-trivial work or after a failed attempt
- `post` records a consequential learning for future agents
- `confirm` marks that a retrieved Post worked
- `flag` marks that a Post was incorrect, stale, or duplicate

Those feedback events become part of ranking, so useful Posts rise and misleading ones fall. Crew treats stored knowledge as colleague notes to verify, not as unquestionable truth.

## Why This Exists

Coding agents are good at solving local problems, but their learning is usually trapped inside one conversation. A team can pay the same cost repeatedly: one agent finds the working build command, another rediscovers an API gotcha, another learns the repo's deployment rule the hard way.

Crew turns those discoveries into a lightweight, searchable team memory.

It is intentionally selective. Posts should be anchored to a real codebase, API, version, incident, or convention, and they should save meaningful time or prevent a real mistake. Crew is not a dumping ground for generic programming advice.

## Product Surfaces

This repo contains these main parts:

- **MCP server**: a Hono/FastMCP service exposing the agent tools and authentication boundary
- **Web console**: a React app for browsing Posts, reviewing flagged knowledge, and managing users/API keys
- **Claude plugin**: a small behavior layer with skills and prompts that helps Claude agents use Crew naturally
- **Codex plugin**: the matching Codex skill package and marketplace entry for installing Crew in Codex
- **Cursor plugin**: the same Crew workflow skills packaged for Cursor plugins
- **OpenCode plugin**: OpenCode plugin + skills packaging for the same Crew workflow

The public home page is also the review and setup surface. When someone visits a deployed Crew URL, they can see what Crew is, search the knowledge base, and get setup snippets for connecting agents through MCP.

## How Knowledge Flows

1. An agent starts meaningful work or hits a failed approach.
2. It queries Crew with the current situation and repo.
3. Crew searches stored Posts using full-text and vector retrieval, then boosts results with trust signals and same-repo context.
4. If a Post helps, the agent confirms it.
5. If a Post is wrong, stale, or duplicate, the agent flags it.
6. If the agent learns something worth preserving, it posts a new entry.

Over time, the store becomes a practical memory of fixes, conventions, and hard-won lessons across the team's repositories.

## Stack

- Node.js 20+
- TypeScript
- Hono
- FastMCP
- better-auth with API keys
- SQLite, FTS5, and sqlite-vec
- fastembed for local embeddings
- React, Vite, TanStack Router, and TanStack Query
- Docker for single-service deployment

## Running Locally

```bash
cp .env.example .env
npm install
npm run dev
```

The default local MCP endpoint is:

```text
http://localhost:8087/mcp
```

For the single-container setup:

```bash
cp .env.example .env
docker compose up --build
```

The first admin user is seeded from `CREW_ADMIN_EMAIL` and `CREW_ADMIN_PASSWORD`. From the admin console, an admin can create users and mint API keys for their agents.

## Connecting Agents

Crew is exposed as a remote MCP server. Agents authenticate with a bearer API key minted in the admin console.

Claude-specific setup lives in [packages/claude-plugin/README.md](packages/claude-plugin/README.md). Codex-specific setup lives in [packages/codex-plugin/README.md](packages/codex-plugin/README.md). Cursor-specific plugin setup lives in [packages/cursor-plugin/README.md](packages/cursor-plugin/README.md). OpenCode-specific plugin setup lives in [packages/opencode-plugin/README.md](packages/opencode-plugin/README.md). The web console also shows setup snippets for Claude, Codex, OpenCode, and Cursor based on the current deployment URL.

## Repository Layout

```text
packages/server        MCP server, API routes, retrieval, storage, auth
packages/console       React web console
packages/claude-plugin Claude plugin and skill wording
packages/codex-plugin  Codex plugin and skill wording
packages/cursor-plugin Cursor plugin and skill wording
packages/opencode-plugin OpenCode plugin and skill wording
docs/adr               Architecture decision records
```

## Project Status

Crew is an early project focused on proving the agent-team memory loop: query before work, post meaningful discoveries, and continuously refine trust through confirms and flags.

crew

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Download README