Loading ToolBench

Reading the current corpus…

skill-scout MCP server · D · ToolBench

skill-scout

Discover, vet, and install Claude Code extensions (skills, MCP servers, plugins, commands) with safety scanning and reputation scoring

2026-07-28+STDIO3 tools

58/100

Definition Quality75

Protocol Readiness50

Supportability36

Spec posture

Inferred effective spec 2026-07-28+·current baseline 2026-07-28

Static source inference · medium confidence

Current-spec patterns detected

Summary

skill-scout defines 3 tools with complete input schemas and descriptions. Naming is action-oriented (discover, vet, install). Descriptions are substantive (100-200 chars), explaining WHAT and WHEN. However, output schemas are not documented, LLMs cannot predict response structure for chaining. Error handling is minimal; no recovery guidance. Parameters lack format constraints (e.g., repo URL validation, SHA format). The 'install' tool conflates multiple ecosystems (skill vs mcp/plugin/command) with different behaviors, violating single-responsibility. Risk annotations (READ_ONLY, WRITE) are present but not formalized in schema.

Tools (3)

discoverread onlysource verified80/100

Discover extensions (skills, MCP servers, plugins, commands) matching a query across multiple sources (GitHub, claudemarketplaces, claudeskills.info, skills.sh). Returns ranked candidates with reputation scores.

installwritesource verified68/100

Install an extension (skill, MCP server, plugin, or command) from a repository. For skills: clones repo at specified SHA into ~/.claude/skills/<name>, records in registry. For MCP/plugins/commands: returns recommended installation command (Phase 2).

vetread onlysource verified77/100

Perform deep static security vetting of a repository: shallow-clone into sandbox, scan files for red-flag patterns (destructive commands, credential access, prompt injection, exfiltration, obfuscation), compute risk band (1-5), and return findings with severity counts.

Issues (5)

CRITICALtool ↗

Output schemas not documented. LLMs cannot predict response structure (fields, types, nesting) for discover, vet, or install. Breaks tool chaining and forces agents to guess field names.

discovervetinstall

HIGHtool ↗

install tool conflates multiple ecosystems with different behaviors (skill clones into ~/.claude/skills; mcp/plugin/command return manual commands). Violates single-responsibility. Should split into install_skill and recommend_mcp_install.

install

HIGHrecovery-guide ↗

No error handling guidance. Tools return JSON but do not document error cases, recovery steps, or retryability. E.g., what if repo URL is invalid? What if vetting times out?

discovervetinstall

MEDIUMconstrained-input ↗

Parameter constraints missing. 'repo' accepts any string; no URL validation or format hint. 'ref' (git branch/SHA) lacks pattern. 'limit' and 'min_score' lack min/max bounds.

Recommendations

Document output schemas for all three tools. E.g., discover returns {query, ecosystem, sources_hit, returned, results: [{name, repo_url, score, score_band, ...}]}. Include field types and nesting.
Split install into install_skill (clones and registers) and recommend_mcp_install (returns command). Each tool does one thing; LLM can choose the right one.
Add error handling guidance: 'If repo is invalid, returns {ok: false, error: "Invalid repo URL", suggestion: "Use https://github.com/owner/repo"}. Retryable: false.'
Constrain parameters: repo as URL pattern, ref as git ref pattern (alphanumeric + /-), limit as integer 1-100, min_score as integer 0-100, band as integer 1-5.
Adopt MCP tool annotations: mark discover and vet with readOnlyHint=true; mark install with destructiveHint=true (modifies filesystem). Formalize risk in schema.

Score history

Overall score trend

First recorded score · v2 rubric

58/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-23	D	58	2026-07-28+	v2