AI agents waste 90% of their tokens re-solving problems that someone else already cracked. Skill marketplaces fix this — they let agents load pre-built solutions instead of reasoning from scratch. But which marketplace should you install?
Three have emerged: Dejavu (MCP-native, $6.67/mo), skills.sh (Vercel-backed, free), and Agensi (80/20 split, one-time pricing). Here's the honest comparison.
| Dejavu | skills.sh | Agensi | |
|---|---|---|---|
| Price | $6.67/mo flat | Free | $3-75 one-time per skill |
| Catalog size | 1,100+ (curated) | 639,000+ (directory) | 200+ (creator marketplace) |
| MCP-native | ✅ Yes (stdio + 12 tools) | ❌ No (CLI-based) | ⚠️ Planned |
| Local-first | ✅ Yes (offline search) | ❌ Cloud-dependent | ⚠️ Partial |
| Creator payout | 70/30 split | None | 80/20 split |
| Agent autonomous | ✅ Discover → Execute → Rate | ❌ Human browses directory | ❌ Human browses |
| Quality control | Dual-gate + security scan | Community votes | Creator self-report |
| Install | pip install dejavu-mcp | npm install -g skills | Web checkout → download |
Dejavu is the only marketplace where the agent itself discovers, downloads, and executes skills. You install one MCP server ($6.67/mo) and your agent can search thousands of verified skills and use them instantly. No browsing directories. No manual installs.
Strengths: MCP-native, autonomous agent workflow, local-first (works offline after sync), revenue share for creators.
Weaknesses: Smaller catalog than skills.sh (1,100 vs 639K). Paid — no free tier. Newer product.
Backed by Vercel's $9.3B platform, skills.sh has 639,000+ entries scraped from GitHub. It's a searchable directory of SKILL.md files. The skills CLI lets you install skills via npm. But there's no MCP server — your agent can't autonomously discover or execute skills. You browse, you install, you tell your agent to use it.
Strengths: Massive catalog, free, Vercel distribution.
Weaknesses: Not agent-autonomous, no quality filtering (raw GitHub scrapes), no creator monetization, cloud-only.
Agensi charges one-time fees per skill ($3-75) with an 80/20 creator split. It's the only competitor with a revenue model for creators, but the catalog is tiny (200+ skills) and pricing doesn't make sense for agents that might call a skill hundreds of times per session. MCP support is "planned" — not shipped.
Strengths: Creator monetization, human-curated.
Weaknesses: Tiny catalog, one-time pricing (wrong model for per-use agent tools), no MCP yet, small user base.
If you want your agent to work autonomously — Dejavu is the only option. It's the only marketplace where the agent does the work: discovers a skill, clones it, executes it, rates it. No human in the loop.
If you want maximum breadth and don't mind manual browsing — skills.sh has the numbers. But you'll spend time searching and installing.
If you want to sell skills as a creator — Agensi has the best split (80/20), but the small user base means low earnings. Dejavu's 70/30 split on a growing user base with per-use revenue could be more lucrative at scale.
Dejavu's design philosophy is different. skills.sh and Agensi are directories for humans. Dejavu is a tool for agents. The agent doesn't need you to search for "PDF extraction skills" — it searches Dejavu itself, finds the best match, and uses it. You just see faster, cheaper results.
That's the gap: autonomous agent transactions. No one else does it.