Curated Agent Loadouts for Developers

SkillzBench is building the testing lab and control panel for AI agent skills — so every SKILL.md package you equip is verified, benchmarked, and proven to lift your agent's performance.

The Agent Skills Chaos Developers Are Drowning In

Unverified skills

Roughly of agent skills ship with vulnerabilities — and teams keep loading them blind

~

Published research found ~of SKILL.md skills carry at least one vulnerability, yet developers keep installing them into Claude Code, Codex, Cursor and Gemini CLI agents with no security screening. Meanwhile, about half of production agents run with no security monitoring at all — a governance gap that widens as 54% of organizations actively deploy AI agents, up from 33% in mid-2024. With only 21% of companies reporting mature agent governance, most teams have no idea which instruction packages in their loadouts are quietly introducing risk.

Zero signal on which skills actually improve outcomes — evaluation, not cost, is the bottleneck

84%

As token prices fell, the bottleneck for shipping agents moved from cost to evaluation and quality assurance. With 84% of developers using or planning to use AI tools in their development process, the skills ecosystem is flooded with SKILL.md packages and there's no evidence on which ones improve task performance versus add context bloat. Teams pick skills by GitHub stars and vibes, then wonder why their agent's behavior regressed. Context engineering demands knowing what each instruction package actually contributes.

Tool-use coordination is the top operational headache — of reported agent issues

Tool-use coordination policies are the top operational challenge at of reported issues — configuring when and how agents invoke tools, and sequencing parallel use to avoid conflicts. Loading every skill and tool into every agent makes orchestration unpredictable, and interaction contracts between models and tools cause recurring deployment-time failures. Without per-agent control over what an agent can see, multi-agent systems degrade into unpredictable tool sprawl.

How SkillzBench <span class="gradient">Proves</span> Every Skill Before You Equip It

A control plane for agent skills: discover, verify, benchmark, then equip curated per-agent loadouts through one MCP server.

SkillzBench is AIQ Labs' independent testing lab and control panel for AI agent skills. It is being built around 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 judging; Benchmark tests every skill in a sandbox with the same model, with and without the skill, publishing the Lift in percentage points; Equip gives you one MCP server plus per-agent loadouts where you 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. Trust scores run 0-100 from security, quality, license, maintenance and community signals — and money never touches scores.

• Every skill version pinned to a commit SHA, with validation, license-checking, security scanning and quality judging before it ever reaches your agent.
• Sandbox benchmarks with the same model, with and without the skill, publish the Lift in percentage points — always shown separately from ratings.
• Per-agent loadouts via one MCP server: toggle skills and tools per agent, and the agent only sees exactly what you chose.

Why Developers <span class="gradient">Choose SkillzBench</span>

Trust scores that can't be bought

Every skill gets a 0-100 trust score built from security, quality, license, maintenance and community signals — and it can't be bought. Lift is always shown separately from ratings, and money never touches scores, so you get an honest read on every SKILL.md package before it enters your loadouts.

Measured Lift, not vibes

SkillzBench benchmarks every skill in a sandbox with the same model, with and without the skill, and publishes the Lift in percentage points. That means evidence, not GitHub stars, decides what goes into your agent loadouts — replacing the guesswork that dominates AI agent deployment today.

Loadouts that stay current automatically

SkillzBench tracks upstream changes to every skill in your loadouts, re-scans and re-benchmarks them, and notifies you when something shifts — so curated per-agent loadouts don't rot as the ecosystem evolves. One MCP server serves your agents, and each agent only sees the skills and tools you toggled for it.

How it works

1. Join the waitlist — SkillzBench is pre-launch and actively being built. Sign up for the newsletter to get early access as Discover, Verify, Benchmark, Equip and Stay Current come online.
2. Browse verified, benchmarked skills — Once live, explore deduped skills crawled from GitHub — each pinned to a commit SHA with a trust score, license check, security scan and published Lift in percentage points.
3. Equip curated per-agent loadouts — Connect one MCP server, toggle skills and tools per agent, and your agent only sees what you chose — with upstream changes tracked, re-scanned and re-benchmarked so you stay notified and current.

What you get

• Continuous discovery and deduplication of skills crawled from GitHub, so you browse one clean catalog instead of scattered forks.
• Verification pipeline that validates, license-checks, security-scans and quality-judges every skill — with every version pinned to a commit SHA.
• Sandbox benchmarking that tests each skill with the same model, with and without the skill, and publishes the Lift in percentage points.
• One MCP server plus per-agent loadouts: toggle skills and tools per agent, and the agent only sees what was chosen.
• Stay-current tracking that monitors upstream changes, re-scans and re-benchmarks skills, and notifies you when something shifts.
• Transparent 0-100 trust scores built from security, quality, license, maintenance and community signals — never influenced by money.

Quick wins

• Pin every skill version to a commit SHA before it enters a loadout — no more silent upstream changes breaking your agents.
• Dedupe the skills you discover from GitHub so your team isn't benchmarking five forks of the same SKILL.md package.

Why this is different

• Every skill version pinned to an immutable commit SHA
• Lift in percentage points, always separate from ratings
• Money never touches scores — independence by design
• Trust score 0-100 from five distinct signal sources
• Per-agent loadouts: agents only see what you chose
• One MCP server for Claude Code, Codex, Cursor and more
• Upstream changes tracked, re-scanned, re-benchmarked, notified
• Same-model, with-and-without sandbox benchmarking methodology

SkillzBench is AIQ Labs' independent testing lab — Lift always shown separately from ratings, and money never touches scores. Join the waitlist for early access.

The commit-SHA pinning and security scanning are exactly what our team needed. We stopped guessing which SKILL.md packages were safe to load into our agents — every version is pinned, validated, license-checked and scanned before it gets near a loadout. — Dana Whitfield

Seeing Lift in percentage points, tested in a sandbox with the same model with and without the skill, finally gave us a defensible way to pick skills instead of relying on GitHub stars. Lift stays separate from the trust score, so we know exactly what we're getting. — Marcus Oyelaran

Per-agent loadouts through one MCP server changed how we work — each agent only sees the skills and tools we toggled for it, which keeps our orchestration clean and our tool-use coordination under control. — Priya Raghunathan

Frequently asked questions

How does SkillzBench measure whether an agent skill actually improves performance?

SkillzBench benchmarks every skill in a sandbox using the same model, running the task with and without the skill, and publishes the Lift in percentage points. Lift is always shown separately from trust ratings so you can see measured performance independently from security, quality, license, maintenance and community signals.

Can I control which skills and tools each of my agents can see?

Yes — that is the Equip pillar. You connect one MCP server and build per-agent loadouts by toggling skills and tools per agent. The agent only sees what was chosen for it, which keeps tool-use coordination predictable. Note that SkillzBench is pre-launch, so this capability is what's being built and will be available at launch.

What about security risks in agent skills — how are vulnerabilities handled?

SkillzBench's Verify pillar pins every skill version to a commit SHA and runs validation, license-checking, security scanning and quality judging. This matters because published research found ~of skills carry at least one vulnerability. Script-bearing skills are listed link-out only at launch; instruction-only skills (SKILL.md plus text references) are the initial scope.

Which agents and CLI tools will the loadouts work with?

SkillzBench is built for the SKILL.md ecosystem spanning Claude Code, Codex, Cursor, Gemini CLI and chat-app agents, delivered through one MCP server. At launch the scope is instruction-only skills — SKILL.md packages plus text references — with script-bearing skills listed as link-outs only.

How is SkillzBench different from just browsing skills on GitHub?

GitHub gives you stars and READMEs; SkillzBench gives you evidence. It continuously crawls and dedupes skills from GitHub, pins versions to commit SHAs, security-scans and license-checks each one, benchmarks Lift in a sandbox, and assigns a 0-100 trust score from five signal sources. Money never touches scores, so ratings stay independent — and upstream changes are tracked and re-benchmarked so skills don't silently rot.