AgentCrew: Multi-Agent System Chat App Supports Multiple Models

What changed

AgentCrew has been released as a chat application built around a multi-agent system designed to facilitate complex tasks through specialization. The system allows users to construct teams of AI agents, each tailored with specific tools and 'personalities' for distinct roles such as architecting, coding, reviewing, researching, and operating web browsers. These agents can hand off work to one another when a task requires a different expertise, with the user maintaining control and orchestration from a single interface. The application supports multiple AI models and incorporates the Model Context Protocol (MCP).

Installation is straightforward across macOS, Linux, and Windows via shell scripts, or through pip for Python environments. AgentCrew offers flexible API key management, supporting subscription-based providers like OpenCode Go (for curated open-source models) and Command Code, as well as pay-as-you-go options from DeepInfra, Together AI, Fireworks, OpenAI, Anthropic, and Google Gemini. It also integrates with GitHub Copilot and allows for custom providers like Ollama and llama.cpp, which can be configured locally for offline use.

AgentCrew can be launched in several modes: a desktop GUI (agentcrew chat), a terminal interface (agentcrew chat --console), a one-shot job mode for CI/CD or batch processing (agentcrew job), and an HTTP API server (agentcrew a2a-server) for integration with other applications. Users can create agents via a command-line interface (agentcrew create-agent) or by defining them manually in TOML configuration files, specifying their names, descriptions, tools, and system prompts.

Supported Tools and Protocols

AgentCrew provides a suite of tools that can be enabled per agent, including code_analysis, file_editing, web_search, fetch_webpage, browser automation, command_execution, memory, clipboard access, adaptive_learning, and voice capabilities. It also supports MCP tools for external integrations. The application communicates using three protocols: Console/GUI for human-to-agent interaction, A2A (Agent-to-Agent) via HTTP+JSON-RPC for connecting multiple AgentCrew instances, and ACP (Agent Communication Protocol) via WebSocket for custom clients and headless control.

Why it matters for builders

This release provides developers with a framework to build sophisticated, multi-agent workflows that can tackle complex projects requiring diverse skill sets. The ability to define specialized agents and orchestrate their collaboration streamlines development processes, from initial design to final code review. Support for a wide array of LLM providers, including local and open-source options, offers significant flexibility in terms of cost, privacy, and model choice.

Practical impact

Developers can leverage AgentCrew to automate multi-step tasks, such as feature development from idea to pull request, by assigning distinct roles to agents (e.g., Architect, Coder, Reviewer). It can also be used for complex data analysis and report generation by chaining agents specialized in data extraction, ratio analysis, trend spotting, risk assessment, and synthesis. The multi-instance agent network capability allows for distributed agent systems, where agents on different machines can collaborate.

Caveats and source limits

The provided source information focuses on the capabilities and setup of AgentCrew. Specific details regarding performance benchmarks, pricing for subscription-based providers beyond their general description, or independent validation of its multi-model support are not detailed. The latest release version mentioned is v0.17.6, but a specific release date for this version is not provided, only a general 'fresh release' status. The source also mentions '5 AI signals, 5 developer signals' without elaborating on what these signals represent.