alenazaharovaux/share/tool-scout

Tool Scout — find the right tools for your tasks

This skill helps discover tools for project tasks. Not everything needs to be built from scratch — a service, MCP server, skill, plugin, AI model, library, or no-code platform might already solve your problem.

It first checks what you already have installed. Then it searches 5 sources: web search, GitHub, MCP server catalogs, awesome-lists, and skill and plugin catalogs. Sources are selected adaptively based on the task type.

Step 0. Configuration (First Run)

Read the config file at ~/.claude/skills/tool-scout/config.md.

If the config file does not exist, run the setup:

Question 1 — Search engine:

Ask the user: "Which web search tool do you have available in Claude Code?"

Options:

• Exa MCP (recommended) — best results; agent_run works well for deep dives
• WebSearch — built-in Claude Code web search (no setup needed, but less precise)
• Other MCP search — if you have a different search MCP server, specify its tool name

Question 2 — GitHub CLI:

Check automatically: gh --version

• If available → github_cli: true
• If not → github_cli: false (GitHub search falls back to web queries with site:github.com)

Question 3 — Language:

Ask the user: "What language should I use for results and communication?" Default: English.

Write answers to ~/.claude/skills/tool-scout/config.md:

search_engine: exa | websearch | other
search_tool_name: mcp__exa__web_search_exa
research_tool_name: mcp__exa__agent_run
github_cli: true | false
language: english

If a tool from the config is missing in the current session — fall back to built-in WebSearch and mention it in the output.

Input

A task or list of subtasks from the user's prompt:

• Free text: "need to build a UI for a dashboard"
• Plan reference: "find tools for tasks 3-5 from the plan"
• Specific subtask: "best way to generate PDF reports"

If a plan file is referenced — read it and extract tasks.

Process

Step 1. Parse input

Identify specific tasks/subtasks from the prompt. If input is a plan file, read it. Formulate each task as a short search phrase.

Step 2. Check what is already installed

Before searching outside, check whether the user already has a solution:

• Skills: ~/.claude/skills/skill-index.md (if it exists), otherwise the folder list in ~/.claude/skills/. Match task keywords against names and descriptions.
• Plugins: claude plugin list.
• MCP servers: claude mcp list.

Show any match as the first line of the output: "You already have X — it does Y". Still run the external search: something noticeably better may exist.

Step 3. Classify by domain

Group tasks by domain:

• UI/design
• Backend/API
• Data/analytics
• Automation/integration
• Content/text
• Infrastructure/deployment
• Other

Step 4. Search (adaptive source selection)

Determine task type and select sources:

Rule: web + GitHub = always. Other sources = when relevant. If unsure — include all.

Source 1: Web search (always)

Established tools query: "best tools for [task] [current year]"

New tools query: "new AI tool [task] launch [current and previous year]"

Library query (when task needs code): "best [language] library for [task] [current year]"

This is more effective than site:npmjs.com — comparison articles provide more context than registry pages.

Always use the current year. Never hardcode a specific year.

Source 2: GitHub (always)

If `github_cli: true`:

• gh search repos "[task]" --sort=stars --limit=5
• gh search repos "[task] tool" --sort=stars --limit=5

If `github_cli: false` (fallback):

• Web search site:github.com [task] tool

Source 3: MCP catalogs (when task involves integration, automation, Claude Code)

If `github_cli: true`:

• gh search repos "mcp server [task]" --sort=stars
• gh search code "[task]" --repo=modelcontextprotocol/servers --filename=README.md

If `github_cli: false`:

• Web search "mcp server [task]" site:github.com

Additionally (always):

• Web search site:smithery.ai [task]

Source 4: Awesome-lists (when looking for curated tool lists)

If `github_cli: true`:

• gh search repos "awesome-[topic]" --sort=stars --limit=3

If `github_cli: false`:

• Web search awesome [topic] github

If a relevant awesome-list is found — read the README (first 200 lines) and extract relevant tools.

Source 5: Claude Code skills and plugins

A skill is a folder with a SKILL.md file. A plugin is a package of skills, commands, hooks, and MCP servers installed from a marketplace.

Official marketplace (local, offline):

• If ~/.claude/plugins/marketplaces/claude-plugins-official/.claude-plugin/marketplace.json exists — search its plugins by task keywords (name, description fields).

If `github_cli: true`:

• gh search code "[task]" --filename=SKILL.md --limit=10 — finds skills by content. Use 2–3 English words ("pdf redact"). Long phrases return nothing; if empty, shorten and retry.
• gh search repos "claude skill [task]" --sort=stars --limit=5

If `github_cli: false`:

• Web search "SKILL.md" [task] site:github.com

Catalogs (web search or fetch):

• anthropics/skills — official Anthropic skills.
• ComposioHQ/awesome-claude-skills, travisvn/awesome-claude-skills, hesreallyhim/awesome-claude-code — curated lists.
• skillsmp.com — search across skills from public GitHub repos.
• claudemarketplaces.com — directory of plugins, skills, and MCP servers ranked by installs.
• aitmpl.com — Claude Code templates: agents, commands, skills, hooks.
• smithery.ai/skills — skills in the Smithery catalog.

Catalogs verified on 2026-10-20. If a catalog does not open — skip it and mention that in the output.

Parallelism

If there are multiple tasks or domains — run queries across different sources in parallel.

Step 5. Deduplication and table

Compile a single table from results across all sources.

Deduplication: if a tool is found in multiple sources — one row, signals aggregated. Found in multiple places = higher confidence (mention in "Why it fits").

Tool types:

• Already installed — found in step 2
• Skill — folder with SKILL.md, copied into ~/.claude/skills/
• Plugin — installed with claude plugin install
• MCP server — can be connected to Claude Code
• SaaS/service — external web service
• AI model — specialized neural network
• Library — npm/pip package
• API — external API for integration
• No-code — visual builder

Maturity:

• Established — 1+ year, has community
• New — recently launched, promising
• Archived — repo archived, no longer maintained (red flag)

Signals column:

• ★ GitHub stars (visible in gh search results, no extra API calls)
• 🔴 Archived — if repo is archived
• ⚠ Code — the skill or MCP server ships scripts that will run on the user's machine

Last commit date is NOT a signal of abandonment — small tools and skills are often stable and don't need updates.

npm/PyPI download counts — only in Deep Dive (requires extra API calls).

Step 6. Output and offer deep dive

Display the table in chat.

After the table — a 2–3 sentence verdict: what to take, why this one, first action. Example: "Take X: it already does Y and was updated this month. First step — read its SKILL.md and copy it into ~/.claude/skills/."

Then ask:

Want to dig deeper? Say "dig into [name]".

Or ask about a specific source: "any MCP servers for this?", "what about skills?", "any awesome-lists?"

Deep Dive (Level 2)

If the user asks to dig deeper into a specific tool, domain, or source:

1. If research_tool_name from the config (Exa agent_run) is available in the session — run it. Otherwise launch an Agent tool (subagent). Prompt:

• Make 5-8 search queries about the tool
• Find: detailed description, usage examples, pricing, limitations, alternatives
• If request targets a specific source ("any skills?") — targeted search via Source 5
• Get npm/PyPI download counts (if applicable)
• For a skill, plugin, or MCP server — read SKILL.md / README and list scripts, network calls, and required keys
• Return a structured report with a source link for every fact

1. Show result in chat:

[Tool name]

What it is: brief description Price: free / freemium / paid (how much) Maturity: when launched, how many users Signals: ★ stars, downloads, found in [sources] Pros: list Cons/limitations: list Security: what runs locally, what access it needs (for skills, plugins, MCP) Alternatives: list How to integrate: skill / plugin / MCP / API / web interface Link: URL

Important

• Communicate in the language specified in config (default: English)
• Freshness is critical — always search for tools from the current and previous year
• Every link and every number (stars, prices) must come from search results in this session. Never write links from memory.
• Do not recommend dead or abandoned tools (archived, no activity for years)
• If a tool is an MCP server, skill, or plugin, say so explicitly (can be connected to Claude Code)
• Read a third-party skill, plugin, or MCP server in full before installing. Tell the user what code they are about to run.
• If the task is trivial and better solved with code — say so directly
• Search engine and GitHub CLI are abstracted — check config.md