Claude Find Solutions

claude-find-solutions

> A Claude Code slash command that finds ready-made projects on the market before you build from scratch. Composes two existing MCP servers into one license-aware, identity-deduped, per-source-ranked answer.

[](LICENSE) [](https://docs.claude.com/en/docs/claude-code) [](https://modelcontextprotocol.io)

---

What it does

You ask: "Has someone built a markdown parser already?"

The skill triggers /find-solutions, which:

1. Calls idea-reality-mcp (GitHub + HN + npm + PyPI + Product Hunt + SO aggregator)
2. Calls mcp-omnisearch GitHub search sorted by stars
3. Enriches the top 5 candidates with license + last-commit via gh api
4. Dedupes by canonical owner/repo so the same project across npm + GitHub merges into one row
5. Applies per-source popularity thresholds (1k stars means different things in Rust vs JS)
6. Returns a ranked table + a verdict: copy-adapt / reuse-as-dependency / pivot / rebuild

License-aware: a repo with NOASSERTION from GitHub's license detector gets demoted from copy-adapt, even if it has 10x the stars of the runner-up. Because copying code with an unrecognized license is a legal mine.

Demo: /find-solutions markdown parser library

Вердикт: copy-adapt · confidence 0.85 · license permissive (MIT)
Top match markdown-it/markdown-it — 21 487★, MIT, last commit 2 days ago.
markedjs/marked (36 842★) demoted: license NOASSERTION, manual review needed.

| # | Repo | Stars | License | Updated | Stack | |---|------|------:|---------|---------|-------| | 1 | markdown-it/markdown-it | 21 487 | MIT | 2026-05-26 | JS, CommonMark | | 2 | markedjs/marked | 36 842 | NOASSERTION | 2026-05-26 | JS, fast | | 3 | erusev/parsedown | 15 032 | MIT | 2026-02-18 | PHP | | 4 | commonmark/commonmark-java | 2 648 | BSD-2 | 2026-05-04 | Java | | 5 | Canop/termimad | 1 183 | MIT | 2025-11-24 | Rust |

Full example in [examples/markdown-parser-library.md](examples/markdown-parser-library.md).

Install

curl -fsSL https://raw.githubusercontent.com/Refusned/claude-find-solutions/main/install.sh | bash

Or manually:

mkdir -p ~/.claude/skills/find-solutions
curl -fsSL https://raw.githubusercontent.com/Refusned/claude-find-solutions/main/SKILL.md \
  -o ~/.claude/skills/find-solutions/SKILL.md

Restart Claude Code. The skill auto-fires on these phrases (English or Russian):

• "find existing solutions for X"
• "what already exists for X"
• "check competition for X"
• "rebuild or reuse"
• "найди готовое решение"
• "что уже есть на рынке"
• "проверь конкурентов"

Or invoke explicitly: /find-solutions

Prerequisites

Two MCP servers must be in ~/.mcp.json:

{
  "mcpServers": {
    "idea-reality": {
      "command": "uvx",
      "args": ["idea-reality-mcp"]
    },
    "mcp-omnisearch": {
      "command": "npx",
      "args": ["-y", "mcp-omnisearch"],
      "env": {
        "GITHUB_API_KEY": ""
      }
    }
  }
}

gh CLI authenticated (gh auth status) for the license-enrichment step.

Why a thin wrapper instead of a new MCP?

This project started as a plan for a from-scratch Python MCP server (solution-finder-mcp) that would aggregate GitHub + 5 package registries. A code review by OpenAI Codex flagged 5 blockers in that plan:

1. Cross-ecosystem verdict thresholds are not defensible (1k stars means different things across npm/PyPI/crates/Maven)
2. Verdict needs per-source normalization
3. Deduplication by raw URL doesn't merge same project across registries (GitHub URL ≠ npm URL ≠ PyPI URL)
4. License punted to future work conflicts with the stated "copy/adapt" use case
5. The whole project may duplicate idea-reality-mcp 80% — a wrapper is more honest

This skill resolves all five inside a 97-line markdown file that calls existing MCP servers. No new code, no parallel maintenance burden. The architectural lessons live in the verdict logic, not in another async HTTP client.

How the verdict works

Per-source thresholds (not cross-ecosystem)

| Source | "Mainstream" threshold | Notes | |--------|------------------------|-------| | GitHub stars | ≥500 mainstream, ≥5 000 canonical | Boring infra often has fewer stars than UI apps | | npm | ≥10 packages matching query | Saturated market signal | | PyPI | Binary (exists or not) | idea-reality-mcp reports counts, not weekly downloads | | crates.io | Existence + count signal | Crates ecosystem is smaller; thresholds shift | | HN mentions | ≥50 in 12 months | Community awareness |

License-aware verdict

| GitHub license | Verdict effect | |----------------|----------------| | MIT, Apache-2.0, BSD-2/3, ISC, 0BSD, Unlicense | copy-adapt allowed | | GPL-, AGPL-, LGPL-* | Downgrade to reuse-as-dependency-only | | null or NOASSERTION | Block confident copy/adapt, surface as reuse-with-caveats | | Other / proprietary | Flag for legal review |

Identity-resolved dedup

Normalize all github.com// references (strip protocol, .git, trailing slash, case). Merge entries with the same owner/repo from different registries into one row. Pick the entry with richer metadata.

Verdict labels

| Label | When | |-------|------| | copy-adapt | Top match canonical AND permissive license AND fresh (≤365d) | | reuse-as-dependency | Top match mainstream OR canonical, any license OK for usage | | pivot | 3+ close matches but none mainstream by source-specific threshold | | rebuild | Fewer than 3 relevant matches, market open |

How it compares

| | Google | ChatGPT | idea-reality-mcp alone | claude-find-solutions | |---|---|---|---|---| | Runs in agent automatically | no | no | yes | yes | | Aggregated GitHub + HN + npm + PyPI | no | no | yes | yes | | Stars-ranked GitHub supplementary | no | no | no | yes | | License-aware verdict | no | no | no | yes | | Identity-resolved dedup across registries | no | no | no | yes | | Per-source thresholds | no | no | no | yes |

Limitations

• Only as good as idea-reality-mcp and mcp-omnisearch underneath. If they drop a registry, this skill loses that source.
• License enrichment makes 5 gh api calls per invocation. Cached lightly within one tool call; not across calls. With gh authenticated this is well within the 5 000/hour budget.
• No PyPI download counts — idea-reality-mcp doesn't expose them; the skill treats PyPI as binary "exists / not."
• Off-target filtering (e.g., excluding PDF tools when the query is markdown parsers) is keyword-based, not semantic. Edge cases will leak through.

Credits

Built on top of:

• mnemox-ai/idea-reality-mcp — the multi-source aggregator
• spences10/mcp-omnisearch — the unified search MCP

The plan critique that shaped this design came from OpenAI Codex acting as a second-opinion reviewer.

License

[MIT](LICENSE) — copy, adapt, redistribute. Attribution appreciated, not required.

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Refusned
• Source: Refusned/claude-find-solutions
• License: MIT

Install and usage instructions live in the source repository linked above.