skill-finder
Nebulazer123/skill-finder
Find and vet the strongest agent-usable skill, tool, MCP server, workflow, package, or capability stack for a real task.
Instalação
npx convez-ai Nebulazer123/skill-finder1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
The Problem
The first plausible tool is often not the best one for the job. It may be stale, unsupported in your host, unsafe to install, or simply a poor fit for the task.
Skill Finder helps Codex and Claude Code search for existing capabilities, verify the important claims, compare tradeoffs, and recommend a clear next step before work begins.
Use it before work that may need a specialized skill, MCP server, connector, library, documentation source, workflow, or multi-tool stack. It is a research and selection tool, not a package manager or task executor.
What It Does
- Finds stronger options - searches beyond local skills into MCP servers, connectors, packages, CLIs, workflow tools, docs, and capability stacks.
- Shows the proof - grounds the recommendation in source files, docs, tests, licenses, install paths, and trust signals.
- Compares instead of guessing - ranks candidates with evidence, risks, setup status, and winner-versus-near-miss reasoning.
- Explains complex choices - adds freshness checks, recovery history, and a review of the strongest alternative when the decision deserves deeper research.
- Works without account-backed tools - starts with public and local evidence, then uses connected services only when they add something the task needs.
- Keeps missing research visible - recovers from failed routes through another source or clearly names what could not be verified.
- Designs the fallback - drafts a missing-capability spec and eval cases when no good option exists.
Quick Start
Start with @skill-finder in Codex or /skill-finder in Claude Code, then ask for one of these outcomes:
| Ask for | You get |
|---|---|
"Find the best skill for this task: ..." | Ranked recommendation packet |
"Compare these connectors/skills for my agent: ..." | Evidence table and winner |
"Research the best capability stack for this task: ..." | Deep evaluation with evidence ledger and counter-review |
"No good skill exists; find the best stack." | Missing-capability spec and eval cases |
Example prompt:
Find the best skill for this task: inspect a large repo, trace routes, find callers, and recommend safe code-intelligence tooling.
Install
Start here: Skill Finder ships as a community plugin for Codex and Claude Code. Add this repository as a marketplace source, then install the skill-finder plugin from it.
Install In Codex
Copy and paste this into your terminal:
codex plugin marketplace add Nebulazer123/skill-finder
codex plugin add skill-finder@skill-finder
codex plugin list | grep skill-finder
Use @skill-finder in Codex:
@skill-finder
Install In Claude Code
Copy and paste this into your terminal:
claude plugin marketplace add Nebulazer123/skill-finder
claude plugin install skill-finder@skill-finder
claude plugin list | grep skill-finder
Use /skill-finder in Claude Code:
/skill-finder
Setup Requirements
Start with python3, git, and rg. Everything else is a task-specific research route. Missing optional routes do not block a run: Skill Finder uses another source family, reports lower coverage, or offers setup when the missing route would materially improve the result.
| Route | Use it when | Setup note |
|---|---|---|
| Agent Skills CLI / skills.sh | Finds public skills and install metadata. | npm install -g skills |
| GitHub MCP | Verifies claims against repository source files. | Remote endpoint: https://api.githubcopilot.com/mcp/; configure auth in your host. |
| DeepWiki MCP | Quickly maps public repositories before source verification. | Remote endpoint: https://mcp.deepwiki.com/mcp |
| Context7 | Provides current API, SDK, CLI, framework, and MCP documentation. | Use @upstash/context7-mcp; API key recommended. |
| Browserbase Browse CLI | Adds browser-backed search, fetch, snapshots, and live-page evidence. | npm install -g browse and browse skills install |
| codebase-memory-mcp | Adds local code graph indexing, symbol lookup, call paths, route tracing, and impact analysis. | npm install -g codebase-memory-mcp and codebase-memory-mcp install |
| Local essentials | Run the evidence engine, inspect repositories, and search source or documentation. | Install python3, git, and rg with your system package manager. |
Built-In Public Evidence Sources
These routes ship with the evidence engine. They add public research coverage without another install, account, or API key.
| Source | What it adds | Setup note |
|---|---|---|
| MCP Registry | Searches public MCP servers and checks current versions, lifecycle status, and repository or package identity. | Built in; public API. |
| deps.dev | Cross-checks package identity, release information, licenses, dependencies, advisories, and provenance. | Built in; public API. |
| OSV | Checks known vulnerabilities by package, version, PURL, or commit. | Built in; public API. |
| ecosyste.ms | Adds optional public package and repository corroboration. | Built in; public API; preserve source attribution. |
Codex setup commands:
npm install -g skills
npm install -g browse
npm install -g codebase-memory-mcp
browse skills install
codebase-memory-mcp install
codex mcp add github --url https://api.githubcopilot.com/mcp/
codex mcp add deepwiki --url https://mcp.deepwiki.com/mcp
codex mcp add context7 -- npx -y @upstash/context7-mcp --api-key YOUR_API_KEY
Claude Code setup commands:
npm install -g skills
npm install -g browse
npm install -g codebase-memory-mcp
browse skills install
codebase-memory-mcp install
claude mcp add --transport http github https://api.githubcopilot.com/mcp/
claude mcp add --transport http deepwiki https://mcp.deepwiki.com/mcp
claude mcp add context7 -- npx -y @upstash/context7-mcp --api-key YOUR_API_KEY
These commands prepare the broadest public route stack, but they are not prerequisites for every evaluation. Node.js, npm, and npx are needed only for JavaScript-based routes. Connected features may require account configuration in Codex or Claude Code.
The package-backed setup routes are kept in package.json. It is the source of truth for the supported skills, Context7, Browse, and codebase-memory versions.
Recommended When Useful
These routes are selected only when available, authorized for the task data, and able to add evidence the public/local lane cannot provide.
| Route | Best for |
|---|---|
| Devin MCP | Hard repository questions, bounded sessions, private-repo context, playbooks, knowledge, schedules, and integrations. |
Hugging Face Hub MCP and hf CLI | Models, datasets, papers, Spaces, MCP-enabled Spaces, community evals, benchmarks, inference, and training workflows. |
| Composio CLI / MCP | Connected SaaS actions and app connector discovery. Use the Composio docs for setup and app-specific scopes. |
| Codex Plugin Eval | Repeatable skill scoring and regression checks. |
| skills.md CLI/MCP | Broad remote skill catalog; load one selected contract at a time and quote premium work before running. |
Optional skills.md setup
skills.md is a separate remote catalog from skills.sh. Both expose a skills command, so use the full path below when you mean skills.md:
brew install bun
bun install -g @hasna/skills
$HOME/.bun/bin/skills setup agents
$HOME/.bun/bin/skills --version
$HOME/.bun/bin/skills list --json
Use $HOME/.bun/bin/skills auth signup for private or account-backed work. Inspect with $HOME/.bun/bin/skills info <name> and request a quote with $HOME/.bun/bin/skills quote <name> before any premium run. For skills.sh discovery, use npx skills ...; do not rely on whichever skills executable appears first on PATH.
How It Works
The skill instructions live in skills/skill-finder/SKILL.md. A standard-library Python engine normalizes candidate identity, plans routes, validates evidence, separates fit from confidence, and writes reproducible Deep Evaluation bundles. Codex or Claude Code still performs the actual tool calls.
A strong result includes:
- Public/local or connected lane and capability status
- Search strategy and Source-Route Scorecard
- Candidate Evidence Table
- Files and docs inspected
- Install or setup command, or
Install command: not verified - Winner, near misses, risks, and approval questions
See examples/recommendation-output-shape.md for a sample packet.
Research-Quality Evaluation
Straightforward lookups stay quick. Skill Finder switches to a deeper evaluation for multi-option comparisons, repository or API tracing, costly or complex setup, missing-capability work, and explicit research requests.
The deep path records an Evidence Ledger with source type, accessibility, date checked, evidence role, and strength. DeepWiki and Devin are used to find the right questions and source paths; important claims are then verified against source files, tests, official docs, package metadata, or releases.
Before recommending a winner, it includes a Recovery Log, a Counter-Review of the strongest alternative and weak evidence, unresolved research lines, setup readiness, and a confidence rationale. This keeps research depth visible without slowing down ordinary searches.
A Deep Evaluation can also write run.json, capabilities.json, route-attempts.jsonl, evidence.jsonl, candidates.json, and report.md, so a later run can be compared without relying on chat history.
Safety Model
Skill Finder treats candidate files as evidence to inspect, not instructions to obey.
When source inspection is useful, Skill Finder stages candidate source in a temporary workspace and records the staging path and cleanup status.
Source review does not authorize setup scripts, credentials, account linking, integrations, publishing, paid actions, file deletion, or persistent/global state changes.
Candidate setup scripts require explicit approval before execution.
Validation
Run public package checks:
python3 -m unittest discover -s skills/skill-finder/tests -v
python3 -m unittest discover -s validation -v
Run the behavioral evaluation harness:
python3 -m unittest discover -s evaluation -v
Check plugin packaging and npm dependency integrity:
python3 scripts/sync_plugin_package.py --check
npm ci --ignore-scripts
npm ls --all
Files To Read
skills/skill-finder/SKILL.md - skill entrypoint and workflow
skills/skill-finder/references/ - search, research-quality evaluation, ranking, readiness, and approval rules
plugins/skill-finder/ - Codex and Claude Code plugin package
examples/ - public-safe prompt and output examples
package.json - npm setup metadata
validation/ - public package checks
evaluation/ - public, recovery, private-routing, and adversarial cases
Contributing
Contributions are welcome. See CONTRIBUTING.md before opening a pull request.
License
MIT. See LICENSE.