crewAIInc / crewAI

What This Repo Claims

Orchestrate teams of AI agents that collaborate on complex tasks through role-based delegation. No cloud dependency required — works with local LLMs via ollama.

50.8k stars. Current version 1.14.4.

The core promise: pip install crewai → crewai create crew → define agents and tasks → run a crew where agents hand off work to each other and produce a final output.

What I Tested

Environment:

• macOS, Apple Silicon
• repoverifiertest isolated user — no prior crewAI install
• Python 3.13
• ollama/llama3.2 — no API key used at any point

Test 1: Install

pip install crewai

Works. Two dependency conflicts on install — not fatal but documented:

• litellm 1.83.14 requires openai==2.24.0, but installed openai 2.36.0
• streamlit 1.38.0 requires rich<14, but installed rich 14.3.4

crewAI overwrites dependencies from other packages. Isolated venvs are essential.

Test 2: CLI availability

crewai --version
zsh: command not found: crewai

CLI installs to /Users/repoverifiertest/Library/Python/3.13/bin/ which is not in PATH by default. Fix:

export PATH="/Users/repoverifiertest/Library/Python/3.13/bin:$PATH"
crewai --version
# crewai, version 1.14.4

Not documented in the README. First-time users will hit this.

Test 3: Scaffold a crew

crewai create crew test_crew

Clean interactive setup. Prompts for LLM provider — selected ollama. Prompts for model — selected ollama/llama3.1. All files created correctly:

• test_crew/src/test_crew/crew.py
• test_crew/src/test_crew/main.py
• test_crew/src/test_crew/config/agents.yaml
• test_crew/src/test_crew/config/tasks.yaml
• test_crew/.env

Test 4: First run — model not found

crewai run
# ERROR: Model llama3.1 not found

The scaffold defaults to ollama/llama3.1 but the locally available model was llama3.2. The setup wizard offers model options from a list without checking what's actually installed in ollama. Fix:

sed -i '' 's/ollama\/llama3.1/ollama\/llama3.2/g' .env

Straightforward fix once you know what's installed. But a first-time user with no prior ollama experience would be stuck.

Test 5: Two-agent crew execution

crewai run

Crew ran successfully end to end:

• Agent 1 (AI LLMs Senior Data Researcher) received the research task and produced output
• Agent 2 (AI LLMs Reporting Analyst) received Agent 1's output as context and expanded it into a structured report
• report.md created in the project root with the final output

The agent handoff worked exactly as claimed. Agent 2 correctly adapted its output based on what Agent 1 returned — including gracefully handling Agent 1's refusal to discuss future technologies (model behavior, not a framework failure).

Findings

Finding 1: Core multi-agent orchestration works

Two agents ran sequentially, context passed correctly between them, final output written to file. The framework does what it claims.

Finding 2: CLI not in PATH after install

crewai command not found until PATH is manually updated. One-line fix but not documented. Affects all macOS users installing without a venv.

Finding 3: Default scaffold model doesn't match local ollama

crewai create crew defaults to ollama/llama3.1 without checking what's installed. If you don't have llama3.1 pulled, the first run fails. Fix is a one-line .env edit but adds friction for new users.

Finding 4: Dependency conflicts on install

Two version conflicts with openai and rich. Not fatal but signals that crewAI is opinionated about its dependency tree. Use an isolated venv.

Finding 5: No API key required

Entire test ran on local ollama with llama3.2. Zero external API calls. The local-first claim holds completely.

What I Did Not Test

• Hierarchical crew process (manager agent delegating to sub-agents)
• CrewAI Flows (event-driven workflows)
• Tool use (web search, file I/O)
• Multi-crew pipelines
• Memory and knowledge features

Verdict: Solid

crewAI delivers on its core claim. Install it, point it at a local ollama model, define two agents and two tasks, and they will collaborate and produce output — with no API key and no cloud dependency. The framework handles agent orchestration, context passing, and output generation correctly out of the box.

Two setup friction points worth knowing: the CLI isn't in PATH after install, and the default scaffold model may not match what you have in ollama. Both are one-line fixes. Neither is a framework failure.

Included in Solution #4: Multi-Agent Orchestration Stack.