Loading ToolBench
Test-driven harness engineering for Python agents: own the loop, capture failures, and gate changes in CI.
Looplet is a harness engineering framework, not an MCP server. The 29 tools are defined across example cartridges (coder, dep_doctor, git_detective) but lack proper MCP registration. Tool definitions are scattered across multiple files with minimal schema documentation. Most tools have basic descriptions (10-50 chars), but parameter schemas are incomplete or missing type information. No error handling guidance, no output schema documentation, and no security considerations visible. The framework appears designed for internal agent testing rather than production MCP deployment.
Execute bash commands in the workspace
Check license compatibility between packages
Check package information from the package registry
Analyze commit message patterns and quality
Get contributor statistics for the repository
Find files that are frequently changed together
Detect dependency files in a project directory
No output schemas documented. Tools return unstructured results with no field definitions, forcing LLMs to parse free-text responses and guess at downstream field names.
Parameter descriptions are minimal (10-35 chars) and lack format/constraint guidance. E.g., 'cmd' for bash has no mention of safety, working directory, or timeout. 'days' and 'top_n' lack min/max bounds.
Destructive tools (bash, write_file, edit_file, worktree) lack confirmation/dry-run patterns and no error recovery guidance. No indication of what happens on failure or how to undo.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 39 | 2026-07-28+ | v2 |
Analyze the directory structure of the repository
Signal that the agent has completed its task
Edit a file by replacing a section
Analyze the age of files in the repository
Identify files with the most changes
Find alternative packages using the host LLM
Inspect git repository information
Find files matching a glob pattern
Search for text in files
List directory contents
Apply multiple edits to a file
Edit a Jupyter notebook
Parse dependencies from a dependency file
Read a text file from the workspace
Get recent commit activity in the repository
Get an overview of the git repository
Delegate a task to a sub-agent
Internal reasoning tool for the agent
Create or manage a todo item
Fetch content from a web URL
Manage git worktrees
Write content to a file in the workspace
No input validation or error handling guidance visible. Tools accept free-form strings (cmd, pattern, path) with no sanitization against injection or path traversal.
Tools like 'think' and 'done' are meta-control tools, not domain tools. They should not be exposed as MCP tools, they belong in agent orchestration logic, not the tool registry.