convez.ai
Agente
karibew

crewkit-cli

karibew/crewkit-cli

Ver no GitHub
InicialPluginScore · de 0 a 100

Manage your team's Claude Code agents with role-based configs, A/B testing, and accuracy tracking. CLI-first.

Instalação

$npx convez-ai karibew/crewkit-cli

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
0
Acompanhando
0
Saúde
62%
Último commit
há 3 dias
Versão
v0.9.19
Issues abertas
1
Contribuidores
2

README

A tradução para o português deste README ainda está na fila.

crewkit

npm version GitHub release Homebrew

Your standards, every session.

Stop reviewing AI-generated inconsistency. crewkit syncs your team's Claude Code agent configurations, tracks what works, and lets you experiment with improvements.

Install

brew install karibew/tap/crewkit

Or via npm, Chocolatey (Windows), or curl:

npm install -g @crewkit/cli
choco install crewkit
curl -fsSL https://crewkit.io/install.sh | sh

Why crewkit?

  • Team consistency - Everyone uses the same agent configs. No more "works on my machine" for AI prompts.
  • Role-based modes - Juniors get coaching mode (agents guide, don't code). Seniors get full autonomy. Same agents, different behaviors.
  • A/B test your prompts - Create experiments, measure outcomes, deploy winners. Data-driven agent improvement.
  • Track what works - See which configs lead to better code, fewer iterations, and faster completion.

Quick Start

# 1. Connect your account
crewkit auth login

# 2. Initialize your project
crewkit init

# 3. Start coding with your team's agents
crewkit code

That's it. Your team's agent configurations are now synced to .claude/agents/ and ready to use.

Features

Shared Agent Configurations

Define agent configurations at the organization level, then customize per-project. Everyone on your team gets the same base agents with project-specific context.

crewkit resources list              # See available agents
crewkit resources show rails-expert # View agent details

Role-Based Agent Modes

Different team members need different agent behaviors:

RoleModeBehavior
JuniorCoachingAgents guide and explain, don't write code directly
IntermediateCollaborativeAgents suggest and implement with explanation
SeniorAutonomousAgents execute with minimal hand-holding

Modes are applied automatically based on your team role.

A/B Testing for Prompts

Test changes to your agent configurations with real usage data. Experiments are created and managed in the dashboard; the CLI can inspect them:

crewkit experiments show <slug>   # View an experiment's details

Session Tracking

Every coding session is tracked for analysis. See which agents perform best, identify patterns, and continuously improve.

Commands

CommandDescription
crewkitShow status overview (auth, org, project, agents)
crewkit codeStart coding session with synced agents
crewkit auth loginConnect your account
crewkit auth logoutSign out
crewkit auth statusCheck authentication status
crewkit initSet up crewkit for current project
crewkit resources listList available agents
crewkit resources show <name>View agent configuration
crewkit experiments show <slug>View experiment details
crewkit org infoShow organization details
crewkit project listList projects
crewkit project infoShow current project details
crewkit playbooks listView playbooks and conventions for your projects
crewkit sessions listList recent sessions (current project by default)
crewkit sessions show <id>Session detail: timing, cost, analysis, tasks, shares
crewkit sessions share <id>Create a share link (--list / --revoke to manage)
crewkit sessions reportTeam/project digest: sessions, cost, top agents, trends
crewkit sessions importImport historical sessions from JSONL
crewkit blueprint <subcommand>AI-powered project planning (create, list, show, tasks, ...)
crewkit updateUpdate crewkit to the latest version
crewkit feedback <msg>Send feedback
crewkit lsp <subcommand>Code intelligence (LSP) for Claude Code
crewkit lsp doctorDiagnose the LSP setup (--json for scripting)
crewkit mcp serveStdio MCP server exposing crewkit's institutional memory
crewkit whoamiShow current user

Run crewkit --help for full command reference.

Code Intelligence (LSP)

crewkit lsp install registers a Claude Code plugin that runs the built-in crewkit language server (crewkit lsp start --stdio) for go-to-definition and symbol search across Rust, TypeScript/JavaScript, and Ruby. Symbols are indexed into .claude/.crewkit-lsp/index.redb per project.

When code intelligence isn't working, run the diagnostic:

crewkit lsp doctor          # human-readable report
crewkit lsp doctor --json   # machine-readable, for scripts/CI

It checks, with a pass/warn/fail marker and a fix-it hint per line:

  • Claude Code CLI availability
  • Plugin installation and enablement (user, project, and local scopes)
  • The plugin's pinned binary path (exists, executable, version matches the running crewkit — stale after updates is the most common silent failure)
  • Conflicting LSP plugins (official rust-analyzer/typescript/ruby LSPs)
  • redb index health (exists, size, openable, file/symbol counts, leftover per-process index files from dead sessions)
  • tree-sitter parser coverage for the languages detected in the repo
  • A stdio handshake smoke test (spawns the server, sends initialize, expects a response, then shuts down cleanly)

Exits non-zero if any check fails. The TUI sidebar's Session section also shows a one-line LSP status (active · N files, plugin off, or failed with a pointer to crewkit lsp doctor).

Institutional Memory (MCP)

crewkit mcp serve runs a stdio MCP server that gives any MCP-capable agent pull access to your project's institutional memory:

ToolWhat it answers
search_project_contextHybrid search over uploaded docs, meeting transcripts, agent emails/Slack, and vectorized past sessions
find_prior_workPast conversations where the team already solved something similar
get_active_workWho's working on what in this project right now
get_conventionsPlaybook conventions for the project's tech stack
get_blueprint_stateActive Blueprint progress and your assigned tasks

Automatic: sessions launched via crewkit code register the server automatically (a generated --mcp-config pointing back at the crewkit binary, with org/project pre-resolved).

Manual (for claude sessions not launched through crewkit, or other MCP clients):

claude mcp add crewkit -- crewkit mcp serve

Or add to your project's .mcp.json:

{
  "mcpServers": {
    "crewkit": {
      "type": "stdio",
      "command": "crewkit",
      "args": ["mcp", "serve"]
    }
  }
}

In the manual path the server resolves your org/project from the git remote (same resolution crewkit code uses). Override with --org / --project flags or CREWKIT_ORG_ID / CREWKIT_PROJECT_ID env vars. Authentication reuses your crewkit auth login credentials; unauthenticated calls return an actionable error instead of failing the session.

Headless Mode

Run crewkit in scripts and CI pipelines:

# Run with a prompt
crewkit code -p "fix the failing tests"

# JSON output for parsing (text | json | stream-json)
crewkit code -p "add input validation" --output-format json

# Quiet mode with timeout
crewkit code -p "refactor auth module" -q --timeout 600

Headless runs get the same observability as interactive sessions: a conversation is created with trigger headless, hook events stream to the platform, and the session is closed with a success/failure outcome when the run finishes (or timeout if the time limit is hit — exit code 4). Progress messages go to stderr so stdout stays parseable with --output-format json.

Headless mode requires authentication and a detectable project — it exits with an error instead of degrading to an untracked run. If your organization has enabled LLM gateway telemetry, headless runs route Claude's Anthropic API traffic through the local gateway automatically (see below).

LLM Gateway

crewkit code --llm-gateway starts a local proxy on 127.0.0.1 and points the spawned Claude Code process at it (ANTHROPIC_BASE_URL is set on the child process only). The gateway captures per-request telemetry — model, token usage (including cache tokens), duration, stop reason, status — and uploads it to the platform in batches.

Enablement rules (all must hold):

  • The organization has LLM gateway telemetry enabled (org setting)
  • The session is not marked --sensitive
  • The gateway was requested: --llm-gateway, or implicitly by headless mode

Privacy:

  • Request/response bodies are uploaded only when the organization's capture mode is full, and are always secret-redacted first. Otherwise only metadata is sent (null bodies).
  • Local logs in ~/.crewkit/logs/gateway-*.log follow the same body rules, are created owner-only (0600), and are pruned after 7 days.
  • Upload failures are logged and dropped (bounded in-memory queue, no retries) so a flaky network never blocks your session.

--llm-gateway-port <PORT> overrides the default ephemeral port. The gateway is skipped under Bedrock/Vertex routing and in no-auth/offline mode.

Session Resume

Continue where you left off:

crewkit code -c                    # Continue most recent session
crewkit code -r                    # Pick a session to resume
crewkit code -r abc123             # Resume specific session
crewkit code -r abc123 --fork-session  # Fork into new session

Platforms

PlatformArchitectureStatus
macOSApple Silicon (arm64)Supported
macOSIntel (x64)Supported
Linuxx64Supported
Windowsx64Supported

Configuration

crewkit auto-detects your project from git remotes and registers it on first crewkit code. For manual configuration:

crewkit init               # Pin the org/project mapping locally
crewkit init --workspace   # Create and register a multi-repo workspace project

crewkit init writes a local .agent/config.yml with your organization and project settings — useful when git detection is ambiguous. It does not register the project with the API; registration happens via crewkit code (or crewkit init --workspace for workspaces).

Global Flags

FlagDescription
--no-colorDisable colored output
--plainUse ASCII-only output (no Unicode)
-h, --helpShow help for any command
-V, --versionShow version

Environment Variables

VariableDescription
CREWKIT_API_URLAPI endpoint (for self-hosted deployments)
CREWKIT_TOKENAccess token for CI/scripts (takes precedence over stored login)
CREWKIT_WEB_URLWeb/dashboard URL override (for self-hosted deployments)
CREWKIT_REPOSITORY_IDExplicit repository ID for crewkit code (skips detection/linking)
CREWKIT_ENVEnvironment tag for error reporting (default: production)
NO_COLORDisable colored output

Links

Development

Prerequisites

ToolVersionInstall
miselatestbrew install mise
justlatestbrew install just
Claude Codelatestnpm install -g @anthropic-ai/claude-code
# macOS — install mise, just, and Docker
brew install mise just

# PostgreSQL via Docker (recommended) — uses docker-compose.yml in repo root
docker compose up -d       # Postgres 17 on :5432 (password: root)

Getting Started

# Clone and install tool versions (Ruby, Node, Rust via .mise.toml)
git clone git@github.com:karibew/crewkit.git && cd crewkit
mise install

# API
cd api && bundle install && bin/rails db:setup && cd ..

# Dashboard
cd dashboard && npm install && cd ..

# CLI
cd cli && just setup && just build && just install-dev

Running

cd api && bin/dev              # Rails API on :3050
cd dashboard && npm run dev    # Next.js dashboard on :3051
crewkit-dev code               # CLI against local API (auto-compiles)

crewkit-dev is an auto-compile dev runner that points to localhost:3050 and builds from source on each invocation. Install it with just install-dev from the cli/ directory. Use it from any project directory:

crewkit-dev code        # Start session against local API
crewkit-dev auth status # Check auth against local API

License

See Terms of Service for license details.

Ver o original no GitHub

Bom para

Relacionados