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, AI model, library, or no-code platform might already solve your problem.

Searches 5 sources: web search, GitHub, MCP server catalogs, awesome-lists, and package registries. 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, supports both web search and code context search
• 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
code_search_tool_name: mcp__exa__get_code_context_exa
github_cli: true | false
language: english

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. Classify by domain

Group tasks by domain:

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

Step 3. Search (adaptive source selection)

Determine task type and select sources:

Task type	Web	GitHub	MCP catalogs	Awesome
SaaS/service	+
Library/package	+	+		+
MCP server	+	+	+
AI tool	+	+		+
Skill/plugin	+	+		+
Unclear	+	+	+	+

Rule: web + GitHub = always. MCP catalogs and awesome = 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.

Parallelism

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

Step 4. 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").

Task	Tool	Type	Maturity	Signals	Why it fits	Link

Tool types:

• 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

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 5. Output and offer deep dive

Display the table in chat. After the table, 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. Launch an Agent tool (subagent) with a detailed 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 GitHub + awesome
   • Get npm/PyPI download counts (if applicable)
   • Return a structured report

2. 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 Alternatives: list How to integrate: 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
• Do not recommend dead or abandoned tools (archived, no activity for years)
• If a tool is an MCP server, say so explicitly (can be connected to Claude Code)
• If the task is trivial and better solved with code — say so directly
• Search engine and GitHub CLI are abstracted — check config.md