Loading ToolBench
An MCP server that gives any agent the ability to discover and use Agent Skills (SKILL.md) — the same skill system used by Claude Code, OpenClaw, nanobot, GitHub Copilot, and OpenAI Codex.
FastSkills MCP server exhibits significant gaps in definition quality. While it provides 3 tools with descriptions, the schemas are largely absent or incomplete, parameter documentation is minimal, and error handling is not evident. Tool names follow basic verb_noun convention (list_, search_, install_), but parameter schemas lack type definitions and descriptions. The descriptions themselves, while present, are generic and don't follow LLM-optimized patterns. Output schemas are undocumented. No evidence of structured error handling, validation guidance, or recovery paths. This is a typical C-grade community server.
Install a skill from the cloud catalog (SkillsMP) into the local skills directory. Downloads and extracts a skill from the cloud catalog, making it available for immediate use. Requires the SKILLSMP_API_KEY environment variable. The skill is placed in the configured skills directory and becomes immediately discoverable via list_skills().
List all available skills with their name, description, and SKILL.md path. Returns every skill found in the configured skills directory. Each entry includes the skill's name, description (from YAML frontmatter), and the full file path to its SKILL.md. To use a skill: call the view tool with the skill's path to read its SKILL.md before starting the task. The file contains best practices and step-by-step instructions the agent should follow.
Search the cloud skill catalog (SkillsMP) by keyword. Searches the SkillsMP cloud catalog for skills matching the query. Requires the SKILLSMP_API_KEY environment variable to be set. Use install_cloud_skill with the returned skill URL to install a skill.
list_skills has empty input schema ({}) with no parameters documented. Input schema MUST define any parameters the tool accepts, or explicitly state it accepts none. Current empty schema is ambiguous.
search_cloud_skills accepts 'query' parameter but schema shows only type 'string' with minimal description. Missing: minLength, maxLength, pattern constraints, examples of valid queries, and guidance on when to use this vs list_skills.
install_cloud_skill accepts 'skill_url' parameter but lacks validation constraints. No guidance on URL format, expected domain, or what happens if URL is invalid/inaccessible. Error handling path is missing.
No output schemas documented for any tool. LLMs cannot infer what list_skills returns (field names, types, structure), what search_cloud_skills returns (result format, pagination), or what install_cloud_skill returns (success indicator, installed path). Forces agent to guess.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 47 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 35 | - | v1 |
install_cloud_skill is a destructive/write operation (downloads and installs code) but has no destructiveHint annotation, no confirmation step, no dry-run mode, and no pre-execution validation guidance. High risk of unintended installations.
Descriptions lack LLM-optimization. 'Search the cloud skill catalog (SkillsMP) by keyword' is vague, does NOT answer: When to use this vs list_skills? What is a 'skill'? What query syntax is supported? What are common failure modes? Descriptions should be 50-200 chars and explicitly guide selection.
No error handling documented. If SKILLSMP_API_KEY is missing, what error is returned? If skill_url is invalid, what guidance is given? If network fails, can the agent retry? No recovery paths documented for any tool.
Tools search_cloud_skills and install_cloud_skill both depend on SKILLSMP_API_KEY environment variable. No guidance in parameter descriptions that this env var must be set, or what error to expect if missing. Undocumented dependency.
No pagination support evident. If search_cloud_skills returns hundreds of skills, no limit/offset parameters shown. Large result sets waste tokens and degrade LLM reasoning. Should document maximum result count and pagination strategy.