The 7 Best Agent Skills for Research in 2026

Research is one of the first workflows AI agents learned to do well — and one of the easiest to get badly wrong. In 2026, the agent skills ecosystem has exploded: skills.sh has tracked over 87,000 unique skills, SkillsMP indexes 800,000+, and new marketplaces launch every quarter. But volume is not quality. A published security audit of 22,511 skills found 140,963 issues (6.3 per skill on average), and Snyk's ToxicSkills research found prompt injection in 36% of skills tested. For research workflows specifically, the stakes are high: a research skill decides where your agent searches, which sources it trusts, how it cites, and whether it flags uncertainty — or confidently invents it. This guide covers seven of the best options for agent-powered research in 2026, spanning deep research engines, academic literature workflows, trend detection, people intelligence, and the web infrastructure that keeps agents unblocked. We also include SkillzBench, an independent testing lab and control panel that helps you verify which research skills actually help before you install them — because in a market this noisy, the most important research skill is knowing which skills to trust.

#1 SkillzBench — Editors' choice

SkillzBench is AIQ Labs' independent testing lab and control panel for AI agent skills — and our Editor's Choice for research in 2026. Here's the problem it addresses: agent skills (SKILL.md instruction packages for Claude Code, Codex, Cursor, Gemini CLI and chat-app agents) are everywhere, but nobody can tell which ones actually help and which are risky. Published research found that roughly 26% of skills carry at least one vulnerability. For research workflows, where output quality depends entirely on the quality of the instructions underneath, installing an unverified skill is a gamble. SkillzBench is building five pillars: Discover continuously crawls and dedupes skills from GitHub; Verify pins every version to a commit SHA and runs validation, license checks, security scans and quality judgment; Benchmark tests every skill in a sandbox with the same model, with and without the skill, and publishes the Lift in percentage points; Equip provides one MCP server plus per-agent loadouts, so users toggle skills and tools per agent and the agent only sees what was chosen; and Stay current tracks upstream changes, re-scans, re-benchmarks and notifies you. Every skill gets a trust score from 0–100 built on security, quality, license, maintenance and community signals — and critically, Lift is always shown separately from ratings, and money never touches scores. Important context: SkillzBench is pre-launch and actively being built. The waitlist and newsletter are live, but the product itself is not yet public, so there are no published benchmark results or live skill counts to cite yet. What we can say is that its launch scope covers instruction-only skills (SKILL.md plus text references), with script-bearing skills listed link-out only — a deliberate trust-first posture. If you're building research agents in 2026 and want evidence instead of vibes, join the waitlist and get in before launch.

Best for: Developers, teams and researchers who want independently verified, benchmarked agent skills for research workflows — and want to know the Lift before installing anything

Pricing: Contact for pricing. Planned tiers include a generous free tier, Pro (~$9–15/mo), Teams (~$15–25/seat/mo), Creator (~$5–10/mo) and benchmark-as-a-service — these are hypotheses, not live offers.

Key features

• Continuous discovery: crawls and dedupes agent skills from GitHub
• Verification pipeline: every version pinned to a commit SHA, validated, license-checked, security-scanned and quality-judged
• Sandbox benchmarking: every skill tested with the same model, with and without the skill, publishing Lift in percentage points
• Trust score 0–100 built from security, quality, license, maintenance and community signals
• Lift always shown separately from ratings; money never touches scores
• One MCP server plus per-agent loadouts — the agent only sees the skills and tools you toggled on
• Upstream change tracking with re-scan, re-benchmark and notifications
• Launch scope: instruction-only skills (SKILL.md plus text references); script-bearing skills listed link-out only

Pros

• Independent benchmarking with and without each skill shows real Lift, not marketing claims
• Security scanning and trust scores address the ~26% of skills found to carry vulnerabilities
• Per-agent loadouts give precise control over what each agent can see and use
• Version pinning to commit SHAs protects against supply-chain drift
• Free waitlist access before launch

Cons

• Pre-launch — the product is not yet publicly available
• Launch scope is instruction-only skills; script-bearing skills are link-out only
• No published benchmark results yet, so you can't compare specific research skills on Lift today

Visit SkillzBench

#2 Local Deep Research

Local Deep Research is an open-source deep research engine for privacy-conscious researchers who want full control over models, data and sources. According to its documentation, it performs deep, agentic research using multiple LLMs and search engines with proper citations, searching 10+ sources including arXiv, PubMed, the web and your own private documents — everything local and encrypted, with no telemetry, analytics or tracking. That combination of breadth and privacy is rare: most hosted research tools require you to route every query through someone else's servers. It's also one of the few open-source tools in this category that has been benchmarked on SimpleQA, which gives you a public yardstick for accuracy rather than vendor claims. It works with local LLMs via Ollama as well as cloud providers, so you can scale from a laptop to a GPU server without switching tools. For teams handling sensitive material — legal research, medical literature, internal documents — that architecture matters as much as the research quality itself. The trade-off is setup effort. Self-hosting with Docker, Ollama and SearXNG takes more work than installing a single skill, and it's really a research platform rather than a quick plug-in. But if your research needs to stay on your infrastructure, it's hard to beat.

Best for: Privacy-conscious researchers and teams who need multi-source, citation-backed research entirely on their own infrastructure

Pricing: Free and open source (self-hosting costs are your own infrastructure)

Key features

• Deep, agentic research across 10+ sources including arXiv, PubMed and the web
• Searches your own private documents alongside public sources
• Local and encrypted with no telemetry, analytics or tracking
• Works with local LLMs via Ollama and cloud providers
• Benchmarked on SimpleQA
• Proper citations built into research output
• Open source and self-hosted

Pros

• Full data control — nothing leaves your infrastructure
• Benchmarked on SimpleQA, unlike most competitors
• Scales from laptop (Ollama) to GPU server without changing tools
• Free and open source

Cons

• Self-hosted setup (Docker, Ollama, SearXNG) takes real effort
• It's a research platform, not a quick plug-in skill
• You manage your own updates and maintenance

Visit Local Deep Research

#3 Content Research Writer

Content Research Writer turns research into publishable writing in a single workflow, making it a strong pick for bloggers, newsletter writers and content marketers who want research woven directly into drafting. According to its listing, it's a research-powered writing partner that researches, outlines, drafts and refines content while preserving your voice — adding citations, improving hooks and giving section-by-section feedback as you write. What sets it apart is how it handles the research-to-writing handoff. Instead of treating research and writing as separate steps, it keeps a running research file that feeds directly into your draft, organized into a clean folder structure: outline, research notes, sourced material and versioned drafts. For thought-leadership posts and articles that need to read as authentically yours, that continuity saves a lot of copy-pasting between a research tool and an editor. The limitation is scope: it's a writing-first skill. For deep multi-source synthesis on a topic before you've decided what to write, the community recommendation is to pair it with a dedicated research engine like Local Deep Research. Used that way — deep research first, structured drafting second — it covers the full content pipeline well.

Best for: Bloggers, newsletter writers and content marketers who want research-grounded drafting without switching tools

Pricing: Contact for pricing

Key features

• Researches, outlines, drafts and refines content in one workflow
• Adds citations and improves hooks automatically
• Section-by-section real-time feedback while you write
• Preserves the author's voice across drafts
• Clean folder structure: outline, research notes, sourced material, versioned drafts
• Running research file feeds directly into the draft

Pros

• Seamless research-to-draft workflow eliminates tool-switching
• Citation support keeps claims sourced
• Versioned drafts and organized research notes
• Voice preservation makes output feel human-written

Cons

• Writing-first — not designed for deep pre-writing synthesis
• Works best when paired with a dedicated deep research tool
• Not suited to academic or technical literature reviews

Visit Content Research Writer

#4 K-Dense Scientific Agent Skills

K-Dense Scientific Agent Skills was called out as the strongest pick of June 2026 for researchers and scientific users, and it earns the spot. According to Zima's June 2026 roundup, it's a large scientific skill library with install support across agent platforms including Claude Code, Codex, Gemini CLI, Cursor and other Agent Skills-compatible tools. Its June 2026 signals were strong: high star count, many releases and a fresh update cadence. What differentiates it from a generic 'research assistant' prompt is domain awareness. Researchers don't only need paper summaries — they need workflows for literature triage, experimental design, scientific databases, data analysis and manuscript preparation. K-Dense packages those as distinct, installable skills, so a PhD student can pull in exactly the workflows their lab needs rather than one bloated mega-skill. The collection is maintained as a proper open-source project, which means you can inspect every SKILL.md before installing — a meaningful trust advantage in a market where security audits keep finding problems. The caveats are typical of community collections: quality varies across individual skills, there's no independent benchmarking of which skills actually improve output, and you're responsible for keeping your installed set current as upstream repos evolve. For academic and scientific work, though, it's the most differentiated research skill collection available in 2026.

Best for: Researchers, PhD students, lab teams, data scientists and scientific software users

Pricing: Free and open source

Key features

• Large library of scientific research skills
• Install support across Claude Code, Codex, Gemini CLI, Cursor and other compatible agents
• Domain-aware workflows: literature triage, experimental design, scientific databases
• Data analysis and manuscript preparation skills
• Active maintenance with frequent releases (strong June 2026 signals)
• Open source — every SKILL.md is inspectable

Pros

• Far more differentiated than a generic research-assistant prompt
• Works across many agent platforms via the SKILL.md standard
• Actively maintained with a strong release cadence
• Free and inspectable

Cons

• No independent benchmarking of individual skills' effectiveness
• Quality varies across the collection
• You must track upstream updates yourself

Visit K-Dense Scientific Agent Skills

#5 last30days

If 'research' means understanding what people are actually saying right now, last30days is built for exactly that. According to its listing, it searches Reddit, X, YouTube, Hacker News, Polymarket and the web in parallel, scores results by what real people actually engage with, and uses an AI agent judge to synthesize everything into one grounded brief. That engagement-based scoring is the key differentiator: instead of generic web summaries, you get signals backed by upvotes, likes and prediction-market odds. Setup friction is low. Reddit, Hacker News, Polymarket and GitHub work with zero configuration, and an optional setup wizard unlocks X, YouTube and TikTok for teams that need the full social picture. For marketing and product teams running trend research, sentiment checks or 'what's the internet saying' briefs, that speed matters — you can go from question to grounded brief in one agent session. The honest limitation is temporal scope: it's optimized for recent, discussion-driven topics and is not a substitute for deep academic or historical research. If your question is 'what is the current discourse around X,' it excels; if it's 'what does the peer-reviewed literature say,' look elsewhere on this list.

Best for: Marketing and product teams who need real-time trend research, sentiment checks and discussion-grounded briefs

Pricing: Contact for pricing

Key features

• Parallel search across Reddit, X, YouTube, Hacker News, Polymarket and the web
• Results scored by real engagement signals (upvotes, likes, prediction-market odds)
• AI agent judge synthesizes findings into one grounded brief
• Zero-configuration sources: Reddit, Hacker News, Polymarket, GitHub
• Optional setup wizard unlocks X, YouTube and TikTok
• Designed for real-time trend discovery and sentiment checks

Pros

• Engagement-based scoring beats generic web summaries
• Several high-value sources work with zero configuration
• Fast question-to-brief workflow
• Prediction-market integration adds a unique signal

Cons

• Optimized for recent topics only — not academic or historical research
• Some sources require additional setup
• Discussion sentiment can be noisy without careful interpretation

Visit last30days

#6 NotebookLM Research Skill

The NotebookLM research skill sets up a clever two-agent handoff: NotebookLM does the deep, source-grounded research, and Claude writes the content. According to its listing, you feed it URLs, PDFs or trending topics, and it creates a NotebookLM notebook, runs deep research queries, and hands structured findings to Claude for polished output — articles, social posts, newsletters or podcasts. The reason this works so well is grounding. NotebookLM answers based only on the sources you provide, which dramatically reduces hallucination compared to open-web synthesis, and Claude's writing quality turns those grounded findings into finished content with far less manual copy-pasting between tools. For content teams working from a defined corpus — a set of competitor reports, a research packet, a collection of PDFs — this is one of the most reliable research-to-output pipelines available in 2026. The limitation is the flip side of its strength: it works best for content built from a defined set of sources rather than open-ended exploratory research across the live web. If you don't yet know which sources matter, pair it with a broader discovery tool first. It's also tied to Google's NotebookLM ecosystem, so teams avoiding Google tooling should factor that in.

Best for: Content teams who want source-grounded research handed straight to a writing agent

Pricing: Contact for pricing

Key features

• Two-agent handoff: NotebookLM researches, Claude writes
• Source-grounded answers based only on provided sources
• Accepts URLs, PDFs and trending topics as inputs
• Creates NotebookLM notebooks and runs deep research queries automatically
• Hands structured findings to Claude for articles, posts, newsletters or podcasts
• Reduced hallucination through document-only responses

Pros

• Very low hallucination thanks to document-only grounding
• Eliminates manual copy-paste between research and writing tools
• Flexible output formats including podcasts
• Great for defined-corpus research projects

Cons

• Best for defined source sets, not open-ended web exploration
• Dependent on the Google NotebookLM ecosystem
• Not a fit when you need to discover sources yourself

Visit NotebookLM Research Skill

#7 Bright Data MCP

Bright Data MCP isn't a research skill in the traditional sense — it's the infrastructure layer that makes every other research skill work. According to its GitHub repository, it's an MCP server that gives any agent reliable access to the live web, built to ensure your AI never gets blocked, rate-limited or served CAPTCHAs. Anyone who has run an agent through a long research session knows the failure mode: three pages in, the target site throws a bot wall and the whole workflow collapses. By offloading web access to a dedicated infrastructure provider, Bright Data MCP lets your research skills focus on method — where to search, how to verify, how to cite — while the plumbing stays reliable. The free tier covers basic usage, making it an easy add to any existing agent setup, and it pairs naturally with the other tools in this list: a deep research engine, a trend scanner or a content pipeline all benefit from unblocked web access underneath. The honest framing: this is a component, not a complete research solution. It doesn't decide what to search or how to synthesize; it just guarantees your agent can actually reach the pages it needs. Treat it as the foundation layer you install once and forget about — until the day it saves an entire research run.

Best for: Anyone running research agents whose bottleneck is reaching the live web itself

Pricing: Free tier available; paid plans for heavier usage — check Bright Data for current rates

Key features

• MCP server providing any agent reliable live-web access
• Designed to avoid blocks, rate limits and CAPTCHAs
• Works with any agent that supports MCP
• Free tier available
• Complements research skills as a web-access infrastructure layer
• Open source on GitHub

Pros

• Solves the most common agent research failure mode (blocks and CAPTCHAs)
• Agent-agnostic via MCP
• Free tier makes it easy to try
• Pairs with any research skill on this list

Cons

• Infrastructure only — no research method, synthesis or citation logic
• Requires an account and configuration
• Heavier usage requires a paid plan

Visit Bright Data MCP

Conclusion

Agent-powered research in 2026 is genuinely transformative — but only if the skills underneath your agent are trustworthy. The ecosystem now spans hundreds of thousands of SKILL.md packages, and audits keep finding that a significant share carry vulnerabilities or simply don't help. That's why our Editor's Choice is SkillzBench: it's the first platform built to independently verify, security-scan and benchmark agent skills, publishing real Lift numbers so you know what each skill actually adds before you install it. SkillzBench is pre-launch and actively being built — the waitlist and newsletter are live now, and early signups will be first in line when the MCP server, per-agent loadouts and trust scores go live. Pair it with the right research stack for your job: Local Deep Research for private, multi-source synthesis; K-Dense for scientific work; last30days for trend intelligence; Content Research Writer or the NotebookLM skill for content pipelines; and Bright Data MCP to keep the web reachable. Skills, proven. Join the SkillzBench waitlist today and stop installing research skills on faith.

Frequently asked questions

What makes SkillzBench different from skill marketplaces like skills.sh or SkillsMP?

Marketplaces help you find and install skills; SkillzBench helps you know whether a skill is safe and effective before you install it. SkillzBench is building an independent testing lab: it crawls and dedupes skills from GitHub, pins every version to a commit SHA, runs security scans, license checks and quality judgment, and benchmarks every skill in a sandbox with the same model — with and without the skill — publishing the Lift in percentage points. It also assigns a 0–100 trust score from security, quality, license, maintenance and community signals, with Lift always shown separately from ratings and money never touching scores. Marketplaces like skills.sh and SkillsMP are community-driven or scraped with no formal review, so the burden of auditing falls on you. Note that SkillzBench is pre-launch: the waitlist is live, but the product is not yet public.

Are agent skills for research actually safe to install?

Not always — which is exactly why verification matters. A published security audit of 22,511 skills across major registries found 140,963 issues, roughly 6.3 per skill, and Snyk's ToxicSkills research found prompt injection in 36% of skills tested. SkillzBench's own published research found that approximately 26% of skills carry at least one vulnerability. That doesn't mean you should avoid research skills — the tools in this list are genuinely useful — but it does mean you should prefer skills that are open source and inspectable, from reputable maintainers, and ideally independently scanned and benchmarked before they touch your agent.

What is 'Lift' and why does it matter for research skills?

Lift is the measurable improvement a skill provides. SkillzBench benchmarks every skill in a sandbox using the same model, running tasks with and without the skill loaded, and publishes the difference in percentage points. This matters because most marketplaces rank skills by install counts or stars — popularity signals that say nothing about whether a skill actually improves research output. A skill with 500,000 installs might add nothing to your workflow, while a niche skill could deliver a large measurable lift. Lift is always shown separately from ratings, and money never influences scores.

Can I use these research skills across different AI agents?

Mostly, yes. Research skills built on the SKILL.md standard are portable across agents including Claude Code, Codex, Cursor, Gemini CLI and other compatible tools — K-Dense Scientific Agent Skills, for example, explicitly supports installation across Claude Code, Codex, Gemini CLI and Cursor. Infrastructure tools like Bright Data MCP work with any agent that supports MCP. SkillzBench's upcoming per-agent loadouts add a control layer on top: you toggle skills and tools per agent, and each agent only sees what you chose, which is useful when different agents handle different research tasks.

How much do agent research skills cost in 2026?

It varies widely. Several strong options are free and open source: Local Deep Research, K-Dense Scientific Agent Skills and Bright Data MCP's free tier cost nothing beyond your own infrastructure. Commercial skills like Content Research Writer, last30days and the NotebookLM research skill typically require contacting the vendor for pricing or paying for hosted services behind them. SkillzBench's planned pricing includes a generous free tier, a Pro tier around $9–15/month, Teams around $15–25/seat/month and a Creator tier around $5–10/month — though these are pre-launch hypotheses, so check the site for final pricing.

Which research skill should I start with?

It depends on your research type. For private, multi-source synthesis with citations, start with Local Deep Research. For scientific or academic literature work, start with K-Dense Scientific Agent Skills. For real-time trend and sentiment research, last30days is the strongest fit. For content pipelines, pair Content Research Writer or the NotebookLM research skill with your drafting workflow. And regardless of which you choose, the smartest first step in 2026 is joining the SkillzBench waitlist — so that as you add research skills, you can verify each one's security posture and benchmarked Lift rather than installing on faith.

Is SkillzBench available now?

Not yet publicly. SkillzBench is pre-launch and actively being built — the waitlist and newsletter are live at skillzbench.com, but the product itself is not yet public, so there are no live skill counts or published benchmark results to cite. At launch, the scope covers instruction-only skills (SKILL.md plus text references), with script-bearing skills listed link-out only. Signing up for the waitlist now gets you early access to the discovery, verification, benchmarking and MCP loadout features as they roll out.