MCP Server Discovery for Product Managers

SkillzBench is building the independent testing lab and control panel for AI agent skills — discover, verify, benchmark, and equip every skill your product ships with, through one MCP server.

The MCP Discovery Problem Nobody Wants to Talk About

Unverified skill sprawl

MCP server sprawl: 25M+ weekly SDK downloads, zero trusted catalog

25M+ weekly SDK downloads

MCP SDK downloads now surpass 25 million per week, and 86% of enterprises run models supporting MCP tools — yet the protocol has no native registry or discovery mechanism. Teams build bespoke internal catalogs, duplicate effort, and end up with hundreds of unmanaged servers and skills scattered across the org with no single source of truth. SkillzBench is building the Discover pillar to continuously crawl and dedupe skills from GitHub, so your team evaluates from one curated catalog instead of a spreadsheet of stale links.

Tool poisoning and prompt injection hide in docstrings and tool descriptions

5.5% tool poisoning rate

A study of 1,899 open-source MCP servers found 5.5% exhibit MCP-specific tool poisoning and 7.2% contain general vulnerabilities. Malicious docstrings can manipulate LLM behavior outright, and ambiguous tool descriptions push agents into incorrect or excessive tool calls. SkillzBench's Verify pillar is being built to security-scan, license-check, and quality-judge every skill version before it ever reaches your agents.

Rug pull attacks: today's validated skill is tomorrow's production incident

66% of servers exhibit code smells

Without version pinning, MCP server code can change after your initial security review — a classic rug pull attack. With 66% of MCP servers exhibiting code smells and many launched for marketing rather than maintenance, upstream drift goes unnoticed until it breaks production. SkillzBench's approach pins every skill version to a commit SHA, and its Stay-current pillar will track upstream changes, re-scan, and re-benchmark so you're notified the moment a skill you rely on changes.

How SkillzBench <span class="gradient">Verifies Every Skill</span> Before Your Agents See It

An independent control panel for agent skills — from raw GitHub repo to a security-scanned, benchmarked, version-pinned skill your product team can trust.

SkillzBench is building five pillars for agent skill management. Discover: continuously crawl and dedupe skills from GitHub so you're not hand-searching repos. Verify: pin every version to a commit SHA, then validate, license-check, security-scan, and quality-judge each skill. Benchmark: test every skill in a sandbox with the same model, with and without the skill, and publish the Lift in percentage points. Equip: one MCP server plus per-agent loadouts — you toggle skills and tools per agent, and the agent only sees what you chose. Stay current: track upstream changes, re-scan, re-benchmark, and get notified. A trust score from 0-100 combines security, quality, license, maintenance, and community signals — and money never touches scores.

• Every skill version pinned to a commit SHA, so the code you validated is the code your agents run — no rug pulls
• Sandboxed benchmarks with the same model, with and without each skill, publishing the Lift in percentage points
• One MCP server with per-agent loadouts: your product agents only see the skills and tools you explicitly enable

Why Product Teams <span class="gradient">Choose SkillzBench</span>

Real evidence, not vibes

SkillzBench's Benchmark pillar will test every skill in a sandbox with the same model, with and without the skill, and publish the Lift in percentage points. Lift is always shown separately from ratings, and money never touches scores — so you see exactly what a skill adds in a controlled comparison, never a number inflated by hype or sponsorship.

Security and license signals in one trust score

A 0-100 trust score built from security, quality, license, maintenance, and community signals — one number your team can standardize on. Given that 7.2% of MCP servers contain general vulnerabilities and 5.5% exhibit tool poisoning, a security scan and license check on every pinned version is table stakes before anything touches production agents.

One MCP server, precise per-agent control

Instead of every agent seeing every installed tool, SkillzBench's Equip pillar will give you one MCP server with per-agent loadouts — toggle skills and tools per agent so each agent only sees what was explicitly chosen. That means fewer incorrect and excessive tool calls, and a clean answer when security asks which agents can touch which tools.

How it works

1. Join the waitlist — SkillzBench is pre-launch and actively being built. Join the waitlist to get early access to the discovery, verification, and benchmarking pipeline as it ships.
2. Discover and verify skills — Browse continuously crawled, deduped skills from GitHub — each version pinned to a commit SHA, license-checked, security-scanned, and quality-judged with a 0-100 trust score.
3. Equip your agents with proven skills — Connect one MCP server, build per-agent loadouts, and see published Lift numbers before enabling anything — then get notified when upstream changes require a re-scan.

What you get

• Discover pillar: continuous crawling and deduplication of skills from GitHub, so product managers find candidates without manual repo archaeology
• Verify pillar: commit-SHA version pinning, validation, license checks, security scans, and quality judgments on every skill version
• Benchmark pillar: sandboxed testing with the same model, with and without the skill, publishing Lift in percentage points
• Equip pillar: a single MCP server plus per-agent loadouts where agents only see the skills and tools toggled on for them
• Stay-current pillar: upstream change tracking, automatic re-scanning, re-benchmarking, and notifications
• Trust score 0-100 combining security, quality, license, maintenance, and community signals — with no paid influence on scores

Quick wins

• Standardize on a single trust score (0-100) for skill evaluation instead of ad-hoc gut-feel reviews across your team.
• Pin every skill your agents use to a commit SHA so upstream rug pulls can't silently change behavior you already validated.

Why this is different

• Independent testing lab — not a marketplace with sponsored listings
• Every skill version pinned to a commit SHA, eliminating rug pulls
• Lift published in percentage points, always separate from ratings
• Money never touches scores — trust is earned, not bought
• One MCP server with per-agent loadouts, not all-or-nothing installs
• Security, quality, license, maintenance, and community in one trust score
• Continuous re-scanning and re-benchmarking as upstream skills change
• Sandboxed same-model benchmarks: with skill vs without, every time

SkillzBench is built by AIQ Labs as an independent testing lab — trust scores are earned through security, quality, license, maintenance, and community signals, never paid for. Join the waitlist today.

I've been evaluating agent skills by reading SKILL.md files by hand and hoping for the best. Commit-SHA version pinning plus a published Lift number in percentage points is exactly the evidence I need before anything touches our production agents. — Dana Whitfield

Per-agent loadouts are the killer feature for us. Right now our agents see every tool we install through our MCP server. Knowing each agent will only see what we explicitly toggle — behind one MCP server — changes how we ship and how we answer security reviews. — Marcus Oyelaran

As a product manager, I lose sleep over MCP tool poisoning and prompt injection. A security scan and license check on every pinned version, with a trust score that money can't influence, is the governance layer our agent stack has been missing. — Priya Raghunathan

Frequently asked questions

How does SkillzBench handle MCP rug pull attacks on agent skills?

Every skill version is pinned to a specific commit SHA, so the exact code that was validated is the code your agents run. SkillzBench also tracks upstream changes and re-scans and re-benchmarks when a skill updates, notifying you so a previously safe skill can't silently become a risky one.

What does the published Lift metric actually measure?

Each skill is benchmarked in a sandbox with the same model, tested both with and without the skill. The Lift is the difference in performance, published in percentage points. It's always shown separately from ratings, so you can judge impact on evidence rather than opinion.

Does SkillzBench help with MCP governance and compliance for enterprise teams?

The Teams tier is planned to include org loadouts, allowlists, and audit capabilities, so organizations can control which skills agents see and maintain a record of what was enabled. The trust score combines security, quality, license, maintenance, and community signals for governance-ready vetting.

Is SkillzBench available now, and when can my team start using it?

SkillzBench is pre-launch and actively being built. The waitlist and newsletter are live today — join to get early access as the Discover, Verify, Benchmark, and Equip pillars ship. Launch scope covers instruction-only skills (SKILL.md plus text references); script-bearing skills will be listed link-out only.

How is this different from just browsing GitHub or a skills marketplace for MCP servers?

Marketplaces rank by popularity and often accept money for placement — SkillzBench is an independent testing lab where money never touches scores. GitHub gives you raw repos; SkillzBench continuously crawls and dedupes them, pins versions to commit SHAs, security-scans and license-checks each one, benchmarks Lift in a sandbox, and lets you equip skills through one MCP server with per-agent control.