crewAI · Health snapshot
Created · run_mwileer43a2f462a1cc4425e · github.com/crewAIInc/crewAI @ main · commit 364b87959014 · model claude-opus-5-5
A five-minute read on how healthy this repository is: seven ratings, the weakest point under each, and the findings a quick read turned up.
crewAI is a Python framework for building multi-agent AI systems: agents with roles and tools are organised into crews and flows that call LLM providers, use memory and knowledge stores, and can be exposed to other agents over A2A or MCP. The repository is a uv workspace of five packages (the core crewai library, crewai-tools, the crewai CLI, crewai-core and crewai-files) plus release tooling; it ships as libraries and a CLI, not as a deployed service.
Overall health: Adequate. Tooling, tests and security controls are well above average for an open-source framework, but every agent run unpickles a file from the working directory that agent tools can write to, and error handling swallows failures in 126 places.
| Rating | Scale | Score |
|---|---|---|
| Engineering standards | 7/10 Good | |
| Security | 6/10 Fair | |
| Configuration and secrets | 6/10 Fair | |
| Data and persistence | 6/10 Fair | |
| Error handling and resilience | 6/10 Fair | |
| Availability and recovery | 5/10 Fair | |
| Performance and scale | 6/10 Fair |
Dependencies run from tests into the five packages. The core library and crewai-core, the CLI and the core library, and crewai-tools and the core library each import one another in both directions, so the package boundaries are not one-way.
modules.cjs, so it matches the tables in this section exactly. Columns are dependency depth: a module sits one column right of everything that calls into it. Line thickness is the number of calls and imports crossing that boundary. The 10 modules under “no edge drawn” take part in none of the edges shown, heaviest first. The accented box is the module with the most files.| Module | Files | What it owns |
|---|---|---|
| lib/crewai/src/crewai | 552 | The framework: agents, crews, tasks, flows, LLM providers, memory, knowledge, events, telemetry, A2A and MCP. |
| lib/crewai-tools/src/crewai_tools | 268 | Tool integrations (search, scraping, browsers, databases, vector stores) and RAG loaders. |
| lib/cli/src/crewai_cli | 97 | The crewai command: scaffolding, run, train, deploy, login, checkpoint and memory TUIs. |
| lib/crewai/tests | 89 | Core tests at the package root: crew, flow, project loading, checkpoints, multimodal. |
| lib/cli/tests | 55 | CLI tests. |
| lib/crewai-tools/tests/tools | 52 | Tool tests, mostly against recorded HTTP cassettes. |
| lib/crewai/tests/llms | 43 | LLM provider tests. |
| lib/crewai/tests/utilities | 41 | Tests for shared utilities. |
| lib/crewai-files/src/crewai_files | 36 | File inputs for multimodal calls: provider uploaders, upload cache and cleanup. |
| lib/crewai-core/src/crewai_core | 30 | Shared core: auth providers, token manager, platform API client, paths and constants. |
| lib/crewai/tests/telemetry | 27 | Telemetry and span tests. |
| lib/crewai/tests/agents | 25 | Agent and agent executor tests. |
| lib/crewai/tests/cli | 23 | An older copy of the CLI tests, still run against crewai-core. |
| lib/crewai-tools/tests/rag | 22 | RAG loader tests. |
| lib/crewai/tests/rag | 20 | Core RAG client tests. |
| lib/crewai-files/tests | 15 | File handling and uploader tests. |
| lib/crewai/tests/memory | 14 | Memory and storage backend contract tests. |
| lib/devtools | 14 | Release tooling: version bump, tag, release and docs checks. |
| lib/crewai/tests/tools | 13 | Core tool layer tests. |
| lib/crewai/tests/mcp | 13 | MCP client and transport tests. |
| lib/crewai/tests/hooks | 12 | LLM and tool hook tests. |
| lib/crewai/tests/skills | 12 | Skills loading and cache tests. |
| lib/crewai-core/tests | 12 | Shared core tests. |
| lib/crewai/tests/experimental | 11 | Experimental evaluation feature tests. |
| lib/crewai-tools/tests | 10 | Tool package tests outside the tool and RAG folders. |
| lib/crewai/tests/events | 9 | Event bus tests. |
| lib/cli/tests/authentication | 9 | CLI login and identity provider tests. |
| lib/crewai/tests/tracing | 8 | Execution tracing tests. |
| scripts | 4 | Maintenance scripts: docs version freezing, model catalog update, evaluation runners. |
| (repo root) | 1 | Shared pytest fixtures and cassette scrubbing (conftest.py). |
| lib/crewai-files | 1 | Package manifest. |
| lib/crewai | 1 | Package manifest. |
| lib/cli | 1 | Package manifest. |
| lib/crewai-core | 1 | Package manifest. |
| lib/crewai-tools | 1 | Package manifest. |
| Entry point kind | Count | Where |
|---|---|---|
crewai CLI commands | 56 | lib/cli/src/crewai_cli/cli.py |
| Release tooling console scripts | 5 | lib/devtools/pyproject.toml |
Standalone scripts run as __main__ | 7 | scripts/, lib/crewai-tools/src/crewai_tools/generate_tool_specs.py |
| Generated project templates (crew and flow) | 3 | lib/cli/src/crewai_cli/templates/ |
| Browser event handlers in the flow visualiser | 18 | lib/crewai/src/crewai/flow/visualization/assets/interactive.js |
Weakest: duplication and dead code
Lint, format, type checks, pre-commit, CodeQL and a vulnerability scan all run in CI (.github/workflows/linter.yml, type-checker.yml, codeql.yml), against 192 groups of duplicated bodies (3,291 removable lines, mostly the CLI test suite copied between lib/cli/tests and lib/crewai/tests/cli) and 168 functions over 100 lines, the longest being setup_listeners at 904 lines in lib/crewai/src/crewai/events/event_listener.py.
Weakest: input validation and injection
Real controls exist (an SSRF-guarding adapter in lib/crewai-tools/src/crewai_tools/security/ssrf_adapter.py, a DML gate on NL2SQL, an opt-in gate on callback deserialisation, signature-verified JWTs), but lib/crewai/src/crewai/agent/core.py:1553 unpickles trained_agents_data.pkl from the working directory on every agent task, the same directory FileWriterTool writes to by default.
Weakest: documentation
Configuration is externalised through environment variables and CLI tokens are Fernet-encrypted in a 0700 directory (lib/crewai-core/src/crewai_core/token_manager.py), but 150 distinct variables are read with no .env.example anywhere, and lib/cli/src/crewai_cli/provider.py:162 overwrites the caller's SSL_CERT_FILE for the whole process.
Weakest: schema evolution
Local SQLite stores use WAL and busy timeouts and checkpoints can be pruned (lib/crewai/src/crewai/state/provider/sqlite_provider.py), but every table is created with CREATE TABLE IF NOT EXISTS and nothing migrates it (lib/crewai/src/crewai/flow/persistence/sqlite.py:113), and agent memory keeps prompts and outputs in plain text with no retention limit (lib/crewai/src/crewai/memory/storage/lancedb_storage.py).
Weakest: surfacing and swallowing
Outbound HTTP calls in the tools carry explicit timeouts and LLM clients retry twice by default (lib/crewai/src/crewai/llms/providers/openai/completion.py:385), but the evidence counts 126 empty or comment-only error handlers, a dozen in lib/cli/src/crewai_cli/crew_run_tui.py alone, so failures often vanish without a trace.
Weakest: detection and alerting
As a library, most availability belongs to the host application, and flow state can be restored from checkpoints (lib/crewai/src/crewai/state/checkpoint_listener.py); but the A2A server it can expose has no health check, and the only metric in the evidence is an event sequence counter in lib/crewai/src/crewai/events/base_events.py.
Weakest: load evidence
Async paths run end to end (kickoff_async in lib/crewai/src/crewai/flow/runtime/__init__.py) and tool results can be cached (lib/crewai/src/crewai/agents/cache/cache_handler.py), but the repository holds no benchmark, load test or stated limit, so throughput under many concurrent crews is untested.
| Severity | Findings | Ratings they fall under |
|---|---|---|
| High | 1 | Security rating |
| Medium | 3 | Error handling and resilience rating, Data and persistence rating |
| Low | 8 | Security rating, Configuration and secrets rating, Engineering standards, Availability and recovery rating, Performance and scale rating |
These counts are a floor, not a total: they are at least the findings a quick read turned up. The full report works through every rating criterion and lists each finding with its evidence.
| Severity | Finding | Rating |
|---|---|---|
| High | Agents unpickle a training file from the working directory on every task, a directory agent tools can write to. | Security rating |
| Medium | Failures are silently swallowed by empty or comment-only error handlers across the CLI and library. | Error handling and resilience rating |
| Medium | Agent memory holds prompts and outputs in plain text with no retention limit. | Data and persistence rating |
The Full Audit and Documentation report of this repository documents: