AgentCrew

Coordinate specialist agents for coding, research, and automation in one workspace.

Stars
★ 226
Last updated
1d ago
License
Apache-2.0
Primary language
Python

At a glance

How it runs
CLIDesktop appSelf-hosted service
Works with
Universal · cross-platformOpenAI API · Claude API
Cost
Free software; you pay for model usage
Setup effort
Medium · a few setup steps
You'll need
Python 3.12+Shell / CLINetwork accessLocal filesystemMCP Server
Typical use
A development team asks an Architect to design a data model, a Coder to implement it, and a Reviewer to inspect security issues.
Not a fit if
  • Users who only need a single conversation without tools or multi-agent orchestration
Website
agentcrew.dev ↗ set on GitHub
Maintenance
Commits in the last 30 dayslast commit 2026-11-20latest release v0.24.1 (2026-11-20)
Source review
61/100 · Some gaps

What does this agent do, and when should you use it?

AgentCrew is a Python 3.12+ multi-agent application available through a desktop GUI, terminal, HTTP API, and CI workflows. Configure agents, tools, and instructions in agents.toml, then switch roles in a conversation or delegate subtasks with @AgentName. Depending on enabled tools and the configured model provider, agents can read and edit files, analyze repositories, search the web, operate a browser, run commands, and connect external tools through MCP. It supports providers including OpenAI, Anthropic, Gemini, GitHub Copilot, ChatGPT Codex, xAI, and DeepInfra, as well as custom local providers such as Ollama and llama.cpp. Run it locally or use its A2A HTTP server to connect AgentCrew instances on different machines.

Create an agent with agentcrew create-agent or define its name, description, tools, and system_prompt in ~/.AgentCrew/agents.toml. In the desktop interface launched by agentcrew chat or terminal interface launched by agentcrew chat --console, select a role with /agent <name> and submit a task; use @<name> <task> to pass a subtask and context to another agent. Enable code_analysis to read and analyze files, file_editing to modify files, web_search and fetch_webpage to gather web content, browser to operate web pages, or command_execution to run commands. Use agentcrew job for one-shot jobs and agentcrew a2a-server for an HTTP+JSON-RPC service; ACP WebSocket is also documented for custom clients.

  1. A development team asks an Architect to design a data model, a Coder to implement it, and a Reviewer to inspect security issues.
  2. An engineer maintaining a repository uses code_analysis to inspect files and structure, then assigns changes to an agent with file_editing enabled.
  3. A researcher has a Researcher agent search and extract several web sources before synthesizing a report.
  4. An operations or QA user enables browser for an agent that must navigate a site, fill forms, and click through a workflow.
  5. A team runs one-shot code review or similar file tasks in CI or batch processing with agentcrew job.
  6. A distributed team connects research and coding agents on separate machines through A2A servers.

Key setup facts

Pulled from the install and usage notes below. The project README has the current values.

Install / run
  • curl -LsSf https://agentcrew.dev/install.sh | bash
  • pip install agentcrew-ai
Environment variables
OPENAI_API_KEY
Config files
~/.AgentCrew/agents.toml
Runtime
Python 3.12+
Install notes cover
Windows · macOS · Linux

How do you install or deploy this agent?

Install with the shell script on macOS or Linux, the PowerShell script on Windows, or pip on any platform. Model use generally requires a provider API key or subscription authentication; local providers can be configured in the config file. The README specifies Python 3.12+.

curl -LsSf https://agentcrew.dev/install.sh | bash
powershell -ExecutionPolicy ByPass -c "irm https://agentcrew.dev/install.ps1 | iex"
pip install agentcrew-ai

Example of setting an OpenAI key:

export OPENAI_API_KEY="sk-proj-..."

How do you use this agent?

Launch the desktop GUI or terminal mode. On first launch, AgentCrew guides you through creating an agent; you can also create one from the command line.

agentcrew chat
agentcrew chat --console
agentcrew create-agent

Define an agent in ~/.AgentCrew/agents.toml, for example:

[[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}."""

Choose agents and delegate work in chat:

/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.

To accept remote connections, start an A2A server:

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

What are this agent's strengths and limitations?

Pros
  • Supports role switching and @AgentName handoffs in one conversation for staged work.
  • Offers a desktop GUI, terminal mode, one-shot jobs, and an A2A HTTP service.
  • Agents can be configured with specific code analysis, file editing, web search, browser, and MCP tools.
  • Documents multiple hosted model providers and custom local providers such as Ollama and llama.cpp.
Limitations
  • Model use generally requires an API key or subscription outside local providers; costs and allowances depend on the provider.
  • Requires Python 3.12 or later.
  • Multi-agent workflows require configuring roles, instructions, and tools, which adds setup compared with a single-assistant chat.
  • File editing, command execution, and browser actions depend on enabling the corresponding tools; their scope follows the configuration.

How does this agent compare with similar options?

The README positions AgentCrew for complex tasks that benefit from multiple perspectives, tool access, and handoffs; a single assistant is a better fit for one focused conversation without tools or multi-agent orchestration. It also supports custom providers such as Ollama, llama.cpp, and LM Studio for users who want to configure local models.

Key facts side by side with the most closely related agents.

Agent Source review Form / cost Stars Updated Language Full support on
AgentCrew This agent 61 · Some gaps CLIFree + model costs ★ 226 1d ago Python OpenAI API · Claude API
Alas Agent Workspace 79 · Good Desktop appFree + model costs ★ 25 today Swift Codex · Claude Code
MassGen 50 · Major gaps CLIFree + model costs ★ 1.2k 5mo ago Python Codex · Claude Code · OpenAI API · Claude API
Polter Terminal Agent Supervisor 47 · Major gaps Desktop appFree + model costs ★ 23 9d ago Zig Claude Code

How does FollowAgents rate this agent?

FollowAgents source review · FARS-2.1
Some gaps
61/ 100 5-point scale 3.1 / 5
This review covers the source as of Nov 11, 2026. The project has shipped 2 stable release(s) since, which have not been re-reviewed. The score and verdict describe the reviewed version only. Recent releases ↓
Trust 12/29
Reliability 8/14
Adaptability 14/18
Convention 13/18
Effectiveness 9/13
Verifiability 5/8
Why each dimension lost points
Trust12 / 29 · 2.1/5

The README lists tools that can be enabled per agent, and the macOS/Linux update flow asks for confirmation, giving some control over permissions and user approval. However, file editing, shell execution, browser, clipboard, MCP, and an A2A service that can bind publicly expose broad local or network effects; the evidence does not show fine-grained controls or a complete authorization model, so least_privilege and external_effects score 1. The README describes multiple model providers, Cursor-hosted inference, and that local xAI logout does not revoke the remote session, providing some data-flow transparency. API keys can be stored in a config file, and the sample shows no encryption, redaction, or key-protection measures, so sensitive_data_handling scores 1. Dependencies are pinned and CodeQL is configured, but no dependency vulnerability response evidence is supplied, so dependency_security scores 1. Updates show release notes and request confirmation, but no recovery or rollback path, so rollback scores 1. The repository name, author field, and license are visible, but publisher identity is unverified and the sample does not establish a responsible maintainer or trusted release path, so source_attribution scores 1.

Reliability8 / 14 · 2.9/5

The README, pyproject.toml, and main.py broadly agree on Python 3.12+, package identity, and supported usage modes. main.py also uses a timeout for update checks and includes version comparison, so self_consistency scores 2. Dependencies are pinned, reducing version drift, but the project has many dependencies and relies on multiple external services; the sample provides no availability or fallback plan, so dependency_availability scores 1. Startup, update checking, plugin initialization, and shutdown paths emit errors or logs. Some exceptions are converted silently to None or allow execution to continue, so failure_messages scores 2.

Adaptability14 / 18 · 3.9/5

The README identifies developer scenarios and provides GUI, terminal, batch, HTTP API, CI, research, coding, and automation examples; it also describes when a single assistant is more suitable, so audience_and_scenarios scores 3. Tool purposes and suggested use cases are documented, with support for custom providers, agents, and MCP. Capability boundaries, security restrictions, and exact per-tool permissions are incomplete, so capability_boundaries scores 2. Commands, configuration examples, and agent handoff syntax are clear, so trigger_precision scores 2. Documentation covers macOS/Linux, Windows, pip, containers, local offline models, and remote terminals, but the sample does not fully specify mode-specific environment constraints, so environment_fit scores 2.

Convention13 / 18 · 3.6/5

The README has clear sections, navigation links, and entry points to configuration, guides, and plugin documentation, so information_architecture scores 3. Install commands cover multiple platforms and explain the Python version and API-key options, but the curl-to-shell and PowerShell download-and-execute instructions lack verification guidance, so install_notes scores 2. AgentCrew, agentcrew-ai, and configuration names are recognizable, but product-name casing varies, so naming_stability scores 2. The README provides many commands, configuration snippets, workflows, and provider-selection guidance, so examples_and_faq scores 3. It discloses xAI subscription and logout limits, billing uncertainty, and Cursor-hosted inference, but limitations for other tools and failure boundaries are incomplete, so known_limitations scores 2. Both LICENSE and project metadata identify Apache-2.0, so license scores 3. pyproject.toml declares a version and main.py checks for updates, but the sample contains no changelog or clear release-note process, so versioning_changelog scores 1. An author name and email are listed, but no maintenance team, support commitment, or clear responsibility is established, so maintenance_responsibility scores 1.

Effectiveness9 / 13 · 3.5/5

The product offers multi-agent collaboration, multiple models, MCP, several interaction modes, and reusable configuration; examples show workflows spanning research and coding, supporting ordinary development use. The sample does not provide evidence of output quality or user outcomes, so output_usability and marginal_value score 2. Multiple providers, local models, and scenario-based tool selection can address different budgets and needs, but users must configure credentials and account for many dependencies and external-service costs; cost-benefit guidance is incomplete, so cost_benefit scores 2.

Verifiability5 / 8 · 3.1/5

README commands and feature claims can be compared with the pyproject.toml entry point and dependencies and with parts of main.py. However, the sample contains only selected source and tests, and many product capabilities are documented rather than shown in implementation, so claim_traceability scores 2. The test files describe regression coverage for specific A2A behaviors, while code shows update confirmation, timeouts, and error handling, but these sources do not cover all of the README's broad claims, so cross_source_corroboration scores 2. The documentation distinguishes feature descriptions from illustrative examples, and the code includes concrete error paths. Some claims about cost, capability, and quality are not supported by verifiable evidence in the sample, so fact_inference_separation scores 2.

Risks and how to mitigate them
  • Review each agent's tool configuration before enabling shell execution, file editing, browser, or MCP. The README's A2A example binds to 0.0.0.0, so verify network access controls when deploying.
  • API keys can be stored in a local JSON config; the sample does not describe encryption or permission hardening. For subscription sign-in, consider provider authorization and billing, and note that local logout does not revoke the remote session.
Evidence confidence: Low Reviewed Nov 11, 2026 Reviewed revision e83a03b79e20 New commits since this review; the score may not cover them
See the full review method →
Maintain this project? Add the FARS badge to your README

The badge shows the current review score and verdict and links back to this page. It refreshes about once an hour after a re-review.

FARS

[![FARS review](https://followagents.com/badge/agentcrew.svg)](https://followagents.com/en/agents/agentcrew)
<a href="https://followagents.com/en/agents/agentcrew"><img src="https://followagents.com/badge/agentcrew.svg" alt="FARS review"></a>

Recent releases

Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.

  1. v0.24.1
    Release notes excerpt
    AgentCrew v0.24.1 This release adds file-backed system prompts for local agents and improves first-startup handling for SQLite memory storage. ✨ Added - Local agents can now load prompt sections from UTF-8 files using system_prompt_ref. Multiple files are wrapped with their configured paths and combined in order. - Prompt files are reread on startup and configuration reload. If any referenced file loads, its content takes precedence over the inline system_prompt; otherwise, the inline prompt remains the fallback. …
    View on GitHub ↗
  2. v0.24.0
    Release notes excerpt
    AgentCrew 0.24.0 This release makes SQLite the default memory backend, adds a safer migration path for existing Chroma data, and improves memory capture, retrieval, and application lifecycle management. Highlights - SQLite and FTS5 are now the default memory backend. Existing Chroma memory can be migrated through an interactive startup prompt or with the new agentcrew-migrate-memory command. The Chroma source is preserved during migration. …
    View on GitHub ↗
  3. v0.23.2
    Release notes excerpt
    AgentCrew 0.23.2 This release adds a new LLM provider and model options, improves context management, and expands Cursor’s integration with MCP tools. Highlights - Added the Refuseless provider, including API-key configuration, reasoning-level support, and built-in GLM models. - Expanded model catalogs with GPT-6.1 Terra for OpenAI Codex, updated GitHub Copilot models, and new DeepSeek, Step, Hy4, and Muse models across supported providers. …
    View on GitHub ↗

FAQ

Do I need a paid API key to use AgentCrew?
Not necessarily. The documentation lists subscription authentication options such as ChatGPT Plus/Pro, GitHub Copilot, and OpenCode Go, and supports configuring local providers such as Ollama and llama.cpp. Costs and allowances depend on the provider or subscription.
Can agents edit files or run commands?
Yes, when the relevant tools are enabled for the agent, such as file_editing or command_execution. The README says command execution has rate limits and audit logs.
Can I use just one agent?
Yes. Select one agent in the desktop or terminal chat interface. Multi-agent handoffs are an available workflow, not a requirement for every task.
Can I connect AgentCrew instances on different machines?
Yes. Run agentcrew a2a-server to provide an HTTP+JSON-RPC service, then add remote_agents entries to connect to other instances.
Is AgentCrew open source?
Yes. The source code is on GitHub (saigontechnology/AgentCrew) under the Apache-2.0 licence.
Is AgentCrew safe to use?
FollowAgents' FARS review (Nov 11, 2026) scored AgentCrew 12/29 on trust and safety. None of the eight safety controls the review checks for was missing outright. The repository has changed since that review.
Is AgentCrew still maintained?
Yes. The latest commit on GitHub is from 2026-11-20, within the last 30 days. The latest release is v0.24.1 (2026-11-20).
Who makes AgentCrew?
AgentCrew is published on GitHub by the account saigontechnology. FollowAgents has not verified which person or company is behind that account.
What is AgentCrew's official website?
The project's GitHub repository lists https://agentcrew.dev as its website. That link is set by the repository owner; the source code itself is on GitHub (saigontechnology/AgentCrew).
Was this page helpful?
View on GitHub ↗ Install ↓

Compare agents like this one

The same FARS review applied across the shortlist this agent qualifies for.

Related concepts