AI agent orchestration platform with multi-agent teams, custom tools, and MCP server integration
CrewForm exposes 11 tools (with 5 duplicates), but critical quality gaps severely limit production readiness. Tool definitions lack comprehensive parameter descriptions, output schemas are undocumented, and error handling is absent. Naming is reasonable (verb_noun pattern), but descriptions are generic (50-100 chars, below the 194-char baseline). No evidence of input validation, permission gates, or recovery guidance. The code_interpreter tool is particularly risky, it accepts arbitrary JavaScript with no sandboxing validation, rate limiting, or timeout documentation. Knowledge_search and a2a_delegate are mentioned but not fully visible in the provided source, making their schemas unverifiable.
Delegate a task to an external A2A-compatible agent. Sends a message and returns the result.
Execute JavaScript code in a sandboxed environment. Returns the result of the last expression or console output.
Check text for grammar, spelling, and style issues. Returns a list of issues with suggestions.
Check text for grammar, spelling, and style issues. Returns a list of issues with suggestions. Supports language auto-detection.
Make an HTTP request to a URL. Returns the response body (truncated to 4000 chars).
Make an HTTP request to a URL. Returns the response body (truncated to 4000 chars).
code_interpreter accepts arbitrary JavaScript with no documented sandboxing, timeout, or resource limits. Descriptions lack execution environment constraints, error modes, or safety guarantees.
Output schemas completely undocumented. No tool specifies what fields are returned, data types, or pagination structure. LLMs cannot plan downstream calls or extract required IDs.
Duplicate tool definitions (web_search, http_request, read_file, grammar_check appear twice). Wastes LLM reasoning cycles and creates ambiguity about which variant to call.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2026-07-28+ | v2 |
Search the knowledge base for relevant information from uploaded documents. Returns the most relevant text passages.
Read the contents of a file from a URL. Returns the file text (truncated to 8000 chars).
Read the contents of a file from a URL. Returns the file text (truncated to 8000 chars).
Search the web for current information. Returns relevant text snippets.
Search the web for current information. Returns relevant text snippets.
No error handling guidance. Tools return truncated responses (4000-8000 chars) but descriptions do not explain truncation behavior, retry logic, or how to handle partial results.
http_request and code_interpreter are destructive/write tools but lack confirmation patterns, dry-run modes, or permission gates. No audit trail or rate limiting documented.