Claude Skills Part 1: 5 Tools Worth Installing in 2026

The phrase "Claude Skills" covers a lot of ground depending on who's using it. For some people it means the official Claude Skills framework — portable SKILL.md instruction files that drop into Claude Code or any compatible agent runtime. For others it means any tool that integrates with Claude via MCP (Model Context Protocol), the Anthropic API, or a Claude-native connector. In practice, the most useful things in this space are a mix: Claude Skills format files, MCP servers that pipe external data into Claude, multi-agent frameworks that use Claude under the hood, and open-source tools you can feed into a Claude conversation. None of these are magic. Each has real setup friction, real cost, and real limitations that marketing copy tends to bury.

This is Part 1 of a series on integrations worth actually installing in 2026. Five tools, five honest assessments. The criterion for inclusion: each one either replaces a paid tool, removes a bottleneck in a real developer or operator workflow, or gives Claude capabilities it does not have by default. Part 2 will cover a different category of integrations — code analysis, document pipelines, and infrastructure tooling.

---

Clay — Lead Research That Connects Directly to Claude

Clay (at clay.com) is a B2B data enrichment and workflow automation platform. Its primary function is letting sales and GTM teams build research pipelines that pull from multiple data providers simultaneously — rather than paying for Apollo, Clearbit, LinkedIn Sales Navigator, and Hunter separately and stitching them together in a spreadsheet. Clay acts as the orchestration layer and charges you wholesale rates for the data you actually consume.

The "100+ data sources" figure in the marketing is on the conservative side. Clay's own documentation and independent reviews from 2026 consistently cite 150+ third-party providers in their data marketplace — including ZoomInfo, Apollo, Clearbit, Hunter, People Data Labs, FullContact, and dozens of niche firmographic and technographic providers. The important nuance: in practice, most production workflows use a small waterfall of three to five providers, not all 150. The broader marketplace is there for edge cases — finding mobile numbers, verifying titles in specific geographies, sourcing founder data from secondary markets.

Best for: GTM teams doing account-based outreach, founders doing their own prospecting, RevOps teams building enrichment pipelines without engineering overhead.

Limitations: Clay is not cheap once you factor in data credits at scale. The credit math is non-trivial — after the March 2026 pricing overhaul, Clay now runs two separate credit pools: Data Credits (for enrichment data) and Actions (for workflow operations). The base plans are Launch at $185/month (2,500 Data Credits, 15,000 Actions) and Growth at $495/month (6,000 Data Credits, 40,000 Actions). Enterprise contracts reportedly run around $30,000/year based on third-party procurement data (vendor claim via Vendr). If you are a solo founder sending 200 cold emails a month, the cost math probably does not pencil.

Cost: Launch $185/mo → Growth $495/mo → Enterprise (custom, ~$30k/yr per Vendr). Data costs are additional, charged at marketplace wholesale rates.

How you connect it to Claude: Clay launched a first-party Claude connector on January 26, 2026. It is not just an MCP integration — it is a native connector inside Claude itself. From Clay's announcement: you authorize the Clay connector from within Claude, and Clay's enrichment infrastructure becomes queryable from the conversation window. You can ask things like "find VP-level RevOps leaders at fintech companies in Bangalore who started in the last six months" and get back verified contact data, company intelligence, and a draft message — without leaving Claude.

If you want the MCP path instead (for Claude Code or programmatic use), the official clay-mcp repo provides the configuration:

{
  "mcpServers": {
    "clay": {
      "url": "https://api.clay.com/v3/mcp"
    }
  }
}

Or via npm:

{
  "mcpServers": {
    "clay-mcp": {
      "command": "npx",
      "args": ["@clayhq/clay-mcp@latest"],
      "env": { "CLAY_API_KEY": "YOUR_CLAY_API_KEY" }
    }
  }
}

New users get 500 free Clay credits when they connect. Generate your API key from your Clay dashboard before setting this up.

---

taste-skill — Stop Claude From Generating Generic Frontend Slop

taste-skill is a Claude Skill format file — a SKILL.md — created by Leon Lin (Leonxlnx/taste-skill on GitHub). The premise is blunt: LLMs left to their own defaults produce frontend interfaces that look like every other LLM-generated interface — the same muted color palettes, the same card grid layouts, the same button styles. taste-skill is a portable instruction set that tells the agent to stop defaulting to templates.

The skill ships with seven variants: taste, redesign, soft, output, minimalist, brutalist, and stitch. Each variant applies a different opinionated aesthetic stance. It works with Claude Code, Cursor, Codex, Gemini CLI, v0, Lovable, and any agent runtime that can load a SKILL.md file. The project's own tasteskill.dev site describes it as "the anti-slop frontend framework for AI agents."

Best for: Frontend-heavy projects where you're using Claude Code or Cursor to generate UI, and the default output keeps looking like a Tailwind template.

Limitations: This is instruction-layer guidance, not a code library. It cannot enforce pixel-perfect design systems, cannot replace a proper design token setup, and cannot compensate for a vague prompt. It shifts the agent's aesthetic defaults, but the output quality still depends on model capability and prompt specificity. The star count — rapidly growing, reaching approximately 20,000+ by late May 2026 and continuing to climb — reflects how widely shared this frustration is, not a complexity-of-implementation score.

A note on the "30,000+ stars" claim in the user brief: the star count is real and growing fast, but the number fluctuates week to week. As of early June 2026, verified sources show it in the 20,000–27,000 range, with some snapshots higher. The user's "30,000+" figure appears to reflect a near-future projection or a brief spike rather than a stable count. Treat the star count as indicative of momentum, not a certification.

Cost: Free. MIT license.

How you connect it to Claude: The canonical install uses npx skills (the vercel-labs/skills runner):

npx skills add https://github.com/Leonxlnx/taste-skill --skill "design-taste-frontend"

This installs the SKILL.md to ~/.claude/skills/design-taste-frontend/SKILL.md (global) or .claude/skills/design-taste-frontend/SKILL.md (project-scoped). You can also copy the SKILL.md manually into any project's .claude/skills/ directory or paste the contents directly into a Claude conversation as system context. The skill auto-loads when Claude Code starts in a directory where it is installed.

---

open_deep_research — Multi-Agent Research Framework Using Claude as Backbone

langchain-ai/open_deep_research is an open-source, configurable multi-agent research framework built on LangGraph. The project, introduced by LangChain in late 2024 and actively developed through 2026, generates long-form research reports by deploying parallel sub-agents — each focused on a specific sub-topic — and synthesizing their outputs via a supervisor agent.

The architecture is worth understanding. Rather than one agent doing sequential searches, open_deep_research spawns independent sub-agents that work concurrently, each with isolated context. A supervisor agent decides when to spawn sub-agents based on query complexity, and a writer agent synthesizes the final report. This matters for research quality: isolated per-agent context means less noise bleed-through between sub-topics. The framework is ranked #6 on the Deep Research Bench Leaderboard (100 PhD-level research tasks across 22 fields), with an overall score of 0.4344 — competitive, not dominant.

Best for: Technical teams who want a self-hosted alternative to cloud research services like Perplexity Pro or OpenAI's Deep Research; anyone doing recurring structured research tasks (competitor analysis, regulatory monitoring, technical literature reviews) where vendor cost-per-query adds up.

Limitations: This is not a drop-in replacement for a research team or a consulting report. It is a configurable starting point. Default output quality depends heavily on which search backend and which model you configure. The local LangGraph server requires some Python environment setup and a LangSmith Studio UI connection, which is an additional friction point. Also: the framework uses Anthropic Claude as its default LLM (claude-3-7-sonnet as planner, claude-3-5-sonnet as writer per the configuration defaults) but you are still calling the Anthropic API — you pay per token. For heavy research workloads this cost accumulates.

Stars: ~10,500–11,600 verified across multiple sources as of June 2026. License: MIT.

Cost: Free to run (self-hosted). You pay for Anthropic API tokens and a Tavily API key (free tier available, paid tier $75/month for heavier use).

How you connect it to Claude: The framework uses Claude through the Anthropic API — it does not install "into" Claude, it calls Claude. Setup:

pip install open-deep-research langgraph-cli python-dotenv
pip install "langgraph-cli[inmem]"

Set environment variables:

export ANTHROPIC_API_KEY=your_key_here
export TAVILY_API_KEY=your_key_here

Start the LangGraph server:

langgraph dev

Then open LangSmith Studio to interact with the research agent via a web UI, or call the local API at http://127.0.0.1:2024 directly. Full MCP compatibility means you can also wire open_deep_research as a tool source in Claude Code or Claude Desktop once the server is running.

---

OpenBB — Open-Source Financial Terminal with MCP-Native Claude Integration

OpenBB (GitHub: OpenBB-finance/OpenBB) is an open-source financial data platform positioned as a free alternative to Bloomberg Terminal. With approximately 68,000 GitHub stars (verified via star-history.com at ~67.3k in May 2026, rising to ~68.4k by early June), it is one of the most-starred financial software projects in the open-source ecosystem.

The Bloomberg Terminal comparison is worth grounding: as of 2026, a single Bloomberg Terminal seat costs $31,980/year, with multi-seat pricing around $28,320/seat/year. OpenBB's platform itself is free. You pay for specific data provider API keys if you want premium data feeds (FactSet, Bloomberg data API, Refinitiv) — but the platform layer, the SDK, and the MCP interface are all open source. The "replaces Bloomberg" framing is overstated — Bloomberg has an irreplaceable data moat for institutional fixed income and derivatives data — but OpenBB genuinely replaces a Bloomberg subscription for equity research, macro data, crypto, and quant analytics workflows.

Best for: Quant analysts, independent researchers, fintech founders, and developers who need structured financial data in Python without a Bloomberg seat. Particularly strong for equity screening, macro data aggregation, and building AI-powered investment workflows.

Limitations: Free data sources (FRED, Yahoo Finance, SEC EDGAR, etc.) are good for many retail and research use cases, but institutional-grade tick data, bond data, and options chains still require paid provider keys. The terminal UI, while functional, is not Bloomberg's workflow — professional traders will find it unfamiliar. Claude integration via MCP is real and works, but you are primarily using Claude to query and interpret the data OpenBB surfaces, not running OpenBB analysis inside Claude autonomously.

Cost: Platform: free and open source. Data: free sources included, paid provider keys optional. OpenBB Workspace (the commercial product) has a paid tier for teams.

How you connect it to Claude: OpenBB ships an official openbb-mcp extension. To connect it to Claude Desktop, add this to your claude_desktop_config.json:

{
  "mcpServers": {
    "openbb-mcp": {
      "command": "uvx",
      "args": [
        "--from", "openbb-mcp-server",
        "--with", "openbb",
        "openbb-mcp",
        "--transport", "stdio"
      ]
    }
  }
}

Requires uvx (part of the uv Python toolchain) installed and on your PATH. Restart Claude Desktop after saving. You can also install the Python SDK for programmatic use:

pip install openbb

Once the MCP server is wired in, you can ask Claude questions like "What is the current P/E ratio and 52-week range for INFY?" or "Show me Indian IT sector revenue growth over the last four quarters" and get back structured data OpenBB pulled from its data providers.

---

exo — Run Large Open-Weight Models Across a Cluster of Local Devices

exo-explore/exo (site: exolabs.net) is an open-source distributed inference framework. Its core capability: link multiple machines — MacBooks, Mac Minis, Linux boxes with NVIDIA GPUs, iPhones, Android devices, Raspberry Pis — into a peer-to-peer cluster that behaves as a single inference endpoint. Exo automatically discovers devices on the local network, determines the optimal model layer split across available hardware, and runs inference without a master-worker bottleneck.

One critical clarification upfront: exo does not run Claude. Anthropic's Claude models are proprietary and not available for local deployment. What exo runs is open-weight models — LLaMA variants, Mistral, DeepSeek V3/R1, Qwen, LlaVA, Kimi K2, and anything loadable from Hugging Face. Exo's API is compatible with the Claude Messages format and the OpenAI Chat Completions format, which means you can point Claude API clients (including Claude Code with the --model flag) at your local exo endpoint — but the model responding will be DeepSeek or LLaMA, not Claude. This is a useful pattern for local inference when you are over API quota or want private inference, but it is not a Claude replacement.

Best for: Developers who have multiple Apple Silicon machines (or NVIDIA GPU boxes) and want to run large open-weight models locally — 70B+ parameters — without paying cloud inference costs. Also useful for teams that need private inference for sensitive workloads where data cannot leave the premises.

Limitations: Setup requires Python 3.12+, and getting stable multi-device inference working requires devices on a low-latency local network (Thunderbolt 5 between Mac Studios is optimal; Wi-Fi works but introduces latency). iOS/Android device support exists but these contribute limited memory to the cluster given RAM constraints. The project is production-capable on Apple Silicon but Linux currently runs on CPU by default, which is slow. You are running open-weight models, not frontier closed models — the capability gap with Claude Opus 4 or GPT-4o is real for complex reasoning tasks.

Stars: ~45,100 verified across multiple sources as of June 2026. License: GPL-3.0.

Cost: Free. You pay for your own hardware and electricity.

How you connect it to Claude (accurately: how you use exo as a local Claude-compatible endpoint):

pip install exo
exo

That starts the node. Add other machines on the same network — exo auto-discovers them. The local API runs at http://localhost:52415. To use it with Claude Code as a local model endpoint:

# In Claude Code settings, point to your local exo endpoint
ANTHROPIC_BASE_URL=http://localhost:52415
ANTHROPIC_API_KEY=any_string  # exo does not validate the key

Or hit it with the standard Anthropic Python SDK, pointed at your local base URL. The model name to pass is the exo model identifier (e.g., llama-3.3-70b, deepseek-v3) — not a Claude model name, because you are not running Claude.

---

How to Stack These Together

These five tools are not competing for the same slot. Here are three concrete combinations worth considering:

GTM research pipeline: Clay (via Claude connector) to pull verified contact data and draft outreach → open_deep_research (via local LangGraph server) to generate deep company intelligence reports → Claude to synthesize and write final messages with Clay-sourced contact context in the same window. The friction point is that Clay credits cost money at scale, so scope your Clay queries to high-priority targets and let open_deep_research handle the free-form research.

Local financial analysis workbench: OpenBB MCP wired into Claude Desktop → ask Claude to pull equity data, macro indicators, and sector comparisons from OpenBB → use Claude to write analysis and flag anomalies. If you are in a bandwidth-constrained or data-sensitive environment and want private inference, you can replace the Claude Desktop endpoint with a local exo cluster running DeepSeek or Qwen — you lose Claude's reasoning quality but gain full data privacy.

Anti-slop frontend sprint: Install taste-skill globally via npx skills add, open Claude Code in your project, and start building. The skill loads automatically and applies its aesthetic constraints. If you hit the Anthropic API rate limit mid-sprint, point Claude Code at a local exo endpoint with a capable open-weight model (LLaMA 3.3 70B handles frontend generation reasonably well). You do not need to change your workflow — just swap the base URL.

---

What's Coming in Part 2

Part 2 will cover the infrastructure and code-intelligence tier: tools that integrate with Claude to handle code review, security scanning, CI/CD pipeline analysis, and documentation generation. The selection will include open-source static analysis tools with MCP servers, AI-native code review frameworks, and at least one integration that handles the specific problem of keeping large codebases coherent as Claude-generated code accumulates. Less marketing, more git diff.

---

Comparison Table

Tool	Category	Open Source?	Stars (June 2026)	Pricing	Claude Integration Path	Best For
Clay	GTM / Lead Research	No	—	$185–$495/mo (data extra)	Native Claude connector + MCP (clay-inc/clay-mcp)	Outbound prospecting, account research
taste-skill	Frontend AI quality	Yes (MIT)	~20–27k	Free	Claude Skills format (npx skills add)	Preventing generic LLM frontend output
open_deep_research	Multi-agent research	Yes (MIT)	~10.5–11.6k	Free (API costs apply)	LangGraph server → Anthropic API	Self-hosted deep research reports
OpenBB	Financial data platform	Yes (AGPL-3.0)	~68k	Free (paid workspace tier)	MCP server (openbb-mcp) → Claude Desktop	Financial analysis replacing Bloomberg
exo	Local LLM inference	Yes (GPL-3.0)	~45.1k	Free (hardware cost)	Claude-format API endpoint (runs open-weight models, NOT Claude)	Private / offline inference on local cluster