AgentCrew

Why AgentCrew?

AgentCrew lets you build a team of specialists, each with the right tools and personality for the job.

• An Architect that weighs trade-offs before a line of code is written.
• A Coder that implements cleanly — no scope creep, no over-engineering.
• A Reviewer that catches what others miss.
• A Researcher that cross-references multiple sources before drawing conclusions.
• A Browser Operator that navigates web apps, fills forms, and clicks through workflows.

Agents hand off work to each other when a task needs a different specialty. You stay in control, orchestrating the team from one interface.

When AgentCrew fits: You have a complex task that benefits from multiple perspectives — building a feature from design to deployment, researching and writing a report, or automating a multi-step workflow.

When a single assistant is enough: You need one focused conversation with no tool access or multi-agent orchestration.

Quick Start

1. Install (10 seconds)

macOS / Linux

curl -LsSf https://agentcrew.dev/install.sh | bash

Windows

powershell -ExecutionPolicy ByPass -c "irm https://agentcrew.dev/install.ps1 | iex"

pip (any platform)

pip install agentcrew-ai

2. Add an API key — or skip this step entirely

Most tools need a paid API key. AgentCrew supports several options that don't require one:

Subscription-based (no API key needed):

# OpenCode Go — curated open-source models (GLM, Kimi, MiMo, Qwen)
agentcrew chat --provider opencode_go

# ChatGPT Plus / Pro (uses Codex models)
agentcrew chatgpt-auth
agentcrew chat --provider openai_codex

# xAI Subscription (supported for eligible accounts)
agentcrew xai-auth
agentcrew chat --provider xai --model-id grok-4.7
# Other curated subscription models
agentcrew chat --provider xai --model-id grok-build-0.1
agentcrew chat --provider xai --model-id grok-4.3
# Local sign-out/reset (does not revoke the remote xAI session)
agentcrew xai-auth --logout

# GitHub Copilot subscribers
agentcrew copilot-auth
agentcrew chat --provider github_copilot

xAI Subscription uses AgentCrew's built-in OAuth device login and the direct xAI Responses API. Availability and model access depend on your current xAI plan and account entitlement. Allowances, overages, and billing are controlled by xAI; a displayed token cost of 0.0 does not mean requests are free. Grok Build 0.1 uses the model's default reasoning because it does not expose selectable reasoning effort through this integration. xai-auth --logout deletes local tokens only; it cannot revoke them remotely.

Pay-as-you-go API keys:

export DEEPINFRA_API_KEY="your-key"       # DeepInfra — open models (LLaMA, Qwen, etc.)
export OPENAI_API_KEY="sk-proj-..."       # OpenAI GPT-4o
export GEMINI_API_KEY="AIza..."           # Google Gemini — free tier available

Or store keys in ~/.AgentCrew/config.json:

{
  "api_keys": {
    "OPENAI_API_KEY": "sk-proj-..."
  }
}

Custom providers:

Custom providers such as Ollama, llama.cpp, and LM Studio are configured in the same ~/.AgentCrew/config.json file. See Custom LLM Providers for the full schema, llama.cpp configuration, model capabilities, and sampling options.

All supported providers:

Which provider should I pick?

If you...	Pick this
Want curated open-source models	OpenCode Go — subscription-based
Want low-cost, open-source friendly	DeepInfra
Have a Command Code subscription	Command Code
Have ChatGPT Plus / Pro	openai_codex — no extra cost
Have GitHub Copilot	github_copilot — no extra cost
Have an eligible xAI subscription	xai — OAuth device login
Want a free tier to start	Google Gemini
Need fully offline / air-gapped	Custom (Ollama, llama.cpp, etc.)

3. Launch

# Desktop GUI (default)
agentcrew chat

# Terminal mode — great for SSH, tmux, low-resource environments
agentcrew chat --console

On the first launch, AgentCrew walks you through creating your first agent.

4. Create your first agent

agentcrew create-agent

Or define one manually in ~/.AgentCrew/agents.toml:

[[agents]]
name = "CodeAssistant"
description = "Helps write and review code"
tools = ["code_analysis", "file_editing", "web_search", "memory"]
system_prompt = """You are an expert software engineer.
Focus on code quality, security, and maintainability.
Today is {current_date}."""

Use Cursor Agent

Cursor Agent is a supported AgentCrew agent type for repository editing and coding tasks. The standard AgentCrew installation includes the Cursor SDK; the Cursor IDE and CLI are not required. Set your Cursor API key, then add a [[cursor_agents]] entry:

export CURSOR_API_KEY="your-cursor-api-key"

[[cursor_agents]]
name = "CursorCoder"
description = "Edits the current repository with Cursor Agent."
model_id = "composer-2.5"
cwd = "."
tools = ["memory"]

Files stay on your machine while inference is hosted by Cursor. Start AgentCrew and select the agent with /agent CursorCoder. See Cursor Agent configuration for all fields and model-selection options.

5. Start working

> /agent Architect
> Design a clean API for a task manager.
>
> @Coding  Implement the task manager in Python using FastAPI.
>
> @Reviewer  Review the code for security issues.

The @AgentName syntax hands off a subtask to another agent. The receiving agent picks up with full context and its own specialized tools and instructions.

What's in the box?

Four ways to use AgentCrew

Job mode example:

agentcrew job --agent "CodeAssistant" \
  "Review for security issues" \
  ./src/**/*.py

A2A server example:

agentcrew a2a-server --host 0.0.0.0 --port 41241

Tools — enable only what each agent needs

Communication protocols

AgentCrew speaks three protocols:

• Console / GUI — Human-to-agent: the chat interface you use daily.
• A2A (Agent-to-Agent) — HTTP+JSON-RPC protocol. Connect multiple AgentCrew instances so one can delegate to agents on another machine.
• ACP (Agent Communication Protocol) — WebSocket protocol for custom clients, IDE integrations, and headless agent control.

Real-world workflows

🧑‍💻 Feature development — from idea to PR

You: @Architect  Design the data model for a multi-tenant SaaS app
     @Coding     Implement it with SQLAlchemy
     @Reviewer   Review the code for security issues

Three agents, one conversation. The Architect thinks about trade-offs, the Coder implements cleanly, the Reviewer catches blind spots.

📊 Financial report generation

# Five specialized agents, each focused on one step
[[agents]]  name = "FinancialDataExtractor"   # parse raw statements
[[agents]]  name = "RatioAnalyst"             # compute key ratios
[[agents]]  name = "TrendAnalyst"             # spot patterns over time
[[agents]]  name = "RiskAssessor"             # flag red flags
[[agents]]  name = "ReportingSynthesizer"     # compile into final report

Each agent handles one step of the pipeline. The Synthesizer collects all findings and writes the final report. See examples/agents/agents-financial-report.toml.

🌐 Multi-instance agent network

# Server A — hosts research agents
server-a$ agentcrew a2a-server --port 41241

# Server B — hosts coding agents
server-b$ agentcrew a2a-server --port 41241

# Your machine — connects to both
# ~/.AgentCrew/agents.toml:
[[remote_agents]]
name = "RemoteResearcher"
url = "http://server-a:41241"

[[remote_agents]]
name = "RemoteCoder"
url = "http://server-b:41241"

Common commands (once you're inside)

Next steps

• GUIDE_GETTING_STARTED.md — Your first 30 minutes with AgentCrew: install, create agents, run a multi-agent workflow.
• GUIDE_WORKFLOWS.md — When and why: pattern guides for common scenarios with decision trees and best practices.
• CONFIGURATION.md — Deep dive into providers, agents, MCP servers, themes, and plugins.
• PLUGIN_DEVELOPMENT.md — Build plugins with EventBus subscriptions and hooks.
• CONTRIBUTING.md — How to build, test, and contribute.
• Docker guide — Running AgentCrew in containers.

License

Apache 2.0 License. See LICENSE for details.