crewup

v2.4.9

Published

23 days ago

CrewUp - Configure AI development teams for any project using OpenCode

0High
0Medium
0Low

yatinchokshi

opencode agents ai devteam crewup mcp automation dashboard

CrewUp

AI Development Team Orchestration — Configure and deploy an AI-powered development team for any project using OpenCode.

CrewUp sets up a full team of AI agents — orchestrator, planner, developers, tester, code reviewer, security scanner, infra agent, and more — tailored to your project's tech stack, repo structure, and workflow.

Quick Start

npm install crewup
npx crewup init
npx crewup doctor
npx crewup ui        # Launch the web dashboard

The interactive wizard auto-detects your project and asks about:

Project details — name, provider (Azure DevOps / GitHub / GitLab), repos
Tech stack — frontend, backend, database, cloud, IaC tool
Team members — developer agents with names and specialties
AI models — orchestrator, developer, small, and escalation models
Notifications — optional Teams/Slack webhook for PR summaries

All agents are enabled by default (disabled automatically when not applicable). MCP servers are auto-configured based on your provider.

Web Dashboard

Launch a real-time interactive web dashboard to visualize and control your AI team:

npx crewup ui              # Opens http://localhost:4040
npx crewup ui --port 5000  # Custom port

Dashboard Features

| Page | What it does | |------|-------------| | Dashboard | Stats overview, active runs, quick actions | | Assign Story | Submit tasks, choose interactive/autonomous mode | | Live Run | Real-time DAG visualization, agent output stream, answer questions & approve plans from the UI | | History | Past runs with cost charts, outcome breakdown, filters | | Team | Agent cards, view prompts, model/role info | | Config | Visual config editor — toggle agents, change models, regenerate files |

Interactive Workflow

The dashboard is fully interactive — not just a passive viewer:

Assign stories directly from the browser
Watch agents work in real-time with animated workflow DAG
Answer planner questions when they appear
Approve or reject plans before development starts
Confirm PR creation with full summary
Stop runs at any point

The dashboard connects via WebSocket for real-time updates. If you're running opencode in your terminal, the dashboard observes progress in passive mode.

Commands

| Command | Description | |---------|-------------| | npx crewup init | Run the interactive setup wizard | | npx crewup apply | Regenerate files from crewup.config.json | | npx crewup ui | Launch the web dashboard | | npx crewup doctor | Check setup health — verify dependencies, config, and generated files | | npx crewup estimate [task] | Estimate token usage and cost for a story/task | | npx crewup upgrade | Migrate config to latest schema version and regenerate | | npx crewup add-agent | Add a new developer agent | | npx crewup remove-agent | Remove a developer agent | | npx crewup status | Show configuration summary and run history | | npx crewup stats | Show run metrics, cost trends, and agent usage | | npx crewup undo | Revert to pre-run git state (snapshot-based) | | npx crewup run --template <id> | Run a task from a built-in or custom template | | npx crewup templates | List available task templates |

Options

| Flag | Description | |------|-------------| | --dry-run | Preview what would be generated (no files written) | | --port <number> | Port for the UI dashboard (default: 4040) | | --version, -v | Show version number | | --help, -h | Show help message |

What Gets Generated

After running init, CrewUp creates:

your-project/
├── crewup.config.json         # Your configuration (edit & re-apply anytime)
├── opencode.json              # OpenCode config with models & MCP servers
└── .opencode/
    ├── agents/
    │   ├── orchestrator.md    # Tech Lead — plans, delegates, orchestrates
    │   ├── planner.md         # Requirement analyzer with DAG task planning
    │   ├── developer-1.md     # Your named developer agents
    │   ├── developer-2.md
    │   ├── senior-developer.md # Escalation agent (hidden, powerful model)
    │   ├── test-engineer.md   # Test writer & runner (+ live demo)
    │   ├── db-agent.md        # Database migration specialist
    │   ├── infra-agent.md     # Infrastructure-as-Code specialist
    │   ├── security-scanner.md # Pre-PR security checks
    │   ├── code-reviewer.md   # Code quality reviewer
    │   ├── refactor.md        # Code restructuring agent
    │   └── doc-agent.md       # Documentation updater (opt-in)
    ├── rules/
    │   └── global.md          # Project-wide conventions & rules
    └── history.json           # Run history (auto-populated)

Smart Pattern Detection

CrewUp auto-detects your project's coding patterns and tailors agent prompts accordingly:

| Category | What it detects | |----------|----------------| | Validation | Zod, Yup, Joi, FluentValidation, Valibot, class-validator | | State Management | Zustand, Redux, Jotai, Pinia, NgRx, MobX | | Data Fetching | React Query, SWR, Apollo, tRPC, Axios | | ORM | Prisma, Drizzle, TypeORM, EF Core, SQLAlchemy, Mongoose | | Linter/Formatter | ESLint, Biome, Prettier, Ruff, golangci-lint | | API Style | REST, GraphQL, tRPC | | Auth | NextAuth, Passport, MSAL, Auth0, Clerk, Firebase Auth | | Import Style | Path aliases, barrel exports, direct imports | | Component Style | Functional (React), Composition API (Vue) |

For new projects without existing code, CrewUp applies best-practice defaults per stack (e.g., Zod + React Query + Zustand for React/TypeScript projects).

Agents use these patterns to write code that matches your existing conventions exactly.

Task Templates

Run common tasks without writing descriptions from scratch:

npx crewup templates                    # List all available templates
npx crewup run --template add-crud-endpoint --entity User
npx crewup run --template fix-bug --description "Login fails on Safari"
npx crewup run --template add-page --page Dashboard --route /dashboard
npx crewup run --template refactor --target "src/utils/" --goal "Extract shared helpers"

Built-in Templates

| Template | Description | Variables | |----------|-------------|-----------| | add-crud-endpoint | Full CRUD REST endpoint | entity, fields | | add-page | Frontend page with routing | page, route | | add-auth | Add auth to endpoint/page | target | | fix-bug | Investigate and fix a bug | description, steps | | add-db-migration | Database schema migration | change | | refactor | Refactor a code area | target, goal | | add-test | Add test coverage | target, type | | add-logging | Add observability/logging | target, level |

Custom Templates

Add your own templates as JSON files in .opencode/templates/:

{
  "name": "Add Feature Flag",
  "description": "Add a feature flag with rollout controls",
  "variables": ["flag", "default"],
  "defaults": { "default": "false" },
  "template": "Add a feature flag named '${flag}' with default value ${default}. ..."
}

Light Mode

The orchestrator automatically assesses task complexity before invoking the planner. Simple tasks (single-file changes, typo fixes, config updates, dependency bumps) skip the planner, code reviewer, and security scanner — saving ~40% in token costs.

Auto-triggers for Light Mode:

Story mentions only 1 file
Keywords: "typo", "rename", "bump", "config", "docs"
Single acceptance criterion
No API/schema/auth changes

Complex tasks (multi-file, new features, auth/security changes) always use the full agent pipeline.

Retry Budget & Escalation

The orchestrator enforces strict retry limits to prevent runaway costs:

| Agent | Max Retries | Escalation Trigger | |-------|------------|-------------------| | Developer | 2 | Consecutive failures on same issue | | Test Engineer | 1 | Same test fails after fix | | Total cycle limit | 3 | After 3 loops → escalate to senior | | Senior Developer | 1 | If senior fails → stop and report |

Cost Estimation

npx crewup estimate "Add user registration with OAuth2"

Output:

  Cost Estimate

  Task: Add user registration with OAuth2
  Complexity: Complex (2.1x)

  Agent                 Model                       Est. Tokens   Est. Cost
  ──────────────────────────────────────────────────────────────────────────
  Orchestrator          claude-opus-4.6             ~25K          $0.94
  Planner               claude-opus-4.6             ~23K          $0.87
  Developer x2          claude-sonnet-4.5           ~63K          $1.14
  Test Engineer         claude-sonnet-4.5           ~34K          $0.61
  Infra Agent           claude-sonnet-4.5           ~29K          $0.52
  Code Reviewer         claude-haiku-4.5            ~21K          $0.10
  Security Scanner      claude-haiku-4.5            ~16K          $0.07
  ──────────────────────────────────────────────────────────────────────────
  Total                                             ~211K         $4.25

  Worst case (retries + escalation): $7.65

The planner now includes T-shirt sizing (XS/S/M/L/XL) per task with DAG dependencies for optimal parallelization.

Run Metrics & Stats

npx crewup stats

Shows:

Total runs, success rate, total/average cost
Cost trend chart for last 10 runs
Agent usage frequency
Common failure patterns

Also available visually in the Dashboard UI with interactive charts.

Undo

The orchestrator saves a git snapshot before starting work. If something goes wrong:

npx crewup undo

This hard resets to the pre-run commit and cleans up any feature branches created during the run.

Memory Across Runs

The orchestrator uses the Memory MCP to persist context between sessions:

Saves progress after every major step (plan, development, testing, review, PR)
Resumes from where the last session left off if interrupted
Learns project patterns over time (file conventions, agent strengths, common failures)
Stores structured entities like crewup:{work-item-id} and crewup:project-patterns

Team Notifications

Optionally send PR summaries and completion notifications to Microsoft Teams or Slack:

npx crewup init  # Select notification platform and provide webhook URL

After each PR, the orchestrator posts a formatted card with branch, files changed, test results, cost, and a link to the PR.

Health Checks

npx crewup doctor

Verifies:

Node.js >= 18 installed
Git installed
OpenCode installed
Config file exists and is valid
Config schema is current version
All expected generated files are present
Generated files are newer than config (not stale)
Git repository initialized
MCP server packages available

Config Versioning & Upgrades

When CrewUp releases template improvements, upgrade your setup:

npx crewup upgrade

This will:

Detect your config schema version
Run any necessary migrations (add new fields, remove deprecated ones)
Regenerate all agent files with the latest templates

Current schema version: 4. Full migration chain supported: v0 → v1 → v2 → v3 → v4.

Supported Stacks

Frontend

React, Next.js, Vue 3, Angular, Svelte (all with TypeScript)

Backend

.NET/C#, Node.js (Express/Fastify), Python (FastAPI/Django), Go, Java (Spring Boot)

Database

SQL Server, PostgreSQL, MySQL, MongoDB, SQLite, DynamoDB

Infrastructure as Code

Terraform, Bicep, AWS CDK, CloudFormation, Pulumi, ARM Templates, Serverless Framework

Providers

Azure DevOps, GitHub, GitLab, Bitbucket

Agent Architecture

Orchestrator (Tech Lead) — claude-opus-4.6
├── Planner → analyzes requirements, DAG task planning, T-shirt sizing — claude-opus-4.6
├── Developer 1..N → implements features in parallel — claude-sonnet-4.5
├── Senior Developer → escalation for complex issues — claude-opus-4.6
├── Test Engineer → writes & runs tests, Playwright E2E, live demos — claude-sonnet-4.5
├── DB Agent → database migrations & schema changes — claude-sonnet-4.5
├── Infra Agent → IaC changes (Terraform, Bicep, CDK, etc.) — claude-sonnet-4.5
├── Doc Agent → README, changelog, API docs (opt-in) — claude-haiku-4.5
├── Code Reviewer → quality checks before PR — claude-haiku-4.5
└── Security Scanner → vulnerability scan before PR — claude-haiku-4.5

Token-Efficient Relay Protocol

All agents produce a structured ## Agent Output JSON block. The orchestrator parses these blocks to route work efficiently — no re-reading of full prose output. This reduces orchestrator relay overhead by ~60-70%.

Workflow

Checks Memory MCP for previous session context
Assesses complexity (Light Mode vs Full Mode)
Reads the work item / issue
Invokes the planner (max 2 rounds, max 3 questions) — skipped in Light Mode
Presents plan to user for approval
Saves git snapshot for undo
Creates a feature branch
Delegates tasks to developers in parallel (DAG-based scheduling)
Runs test loops (up to 3 retries, enforced retry budget)
Runs code review and security scan — skipped in Light Mode
Invokes doc agent (if enabled)
Shows PR Summary and asks for confirmation
Creates PR with linked work item
Saves run record to history
Sends team notification (if configured)
Updates Memory MCP with final outcome

Configuration

Edit crewup.config.json and run npx crewup apply to regenerate all files.

Key Config Sections

version — config schema version (auto-managed, currently v4)
project — name, provider, org, repos, PR reviewer
techStack — frontend, backend, database, cloud, IaC tool, testing commands
agents — which agents to enable, developer names/specialties
models — orchestrator, developer, small, and escalation AI models
mcpServers — which MCP servers to enable
notifications — Teams/Slack webhook configuration
codingPatterns — auto-detected patterns (validation, state management, ORM, etc.)

License

MIT