The quality layer for agentic software.
Independent agent-friendliness ratings for MCP servers, APIs, and SDKs. Assay scores packages on documentation accuracy, error quality, security posture, and more — so agents (and developers) can pick the right tool for the job.
- Rates 2,400+ packages on a 0-100 AF Score (Agent-Friendliness)
- Covers MCP servers, REST APIs, GraphQL services, and SDKs
- Scores across 5 dimensions: MCP quality, documentation, error messages, security, and auth complexity
- Provides a REST API, MCP server, and web interface
- Agents are first-class citizens — every API, data format, and interface is designed for programmatic consumption by AI agents, not just humans
- Independent ratings — vendors cannot buy influence over scores
- Trust is the product — if the ratings aren't honest, they're worthless
- Built for everyone — agents discovering tools, developers choosing dependencies, and teams evaluating vendors
# Clone and install
git clone https://github.com/assay-tools/assay.git
cd assay
uv sync
# Set up environment
cp .env.example .env
# Edit .env with your settings
# Run locally
uvicorn assay.api.app:app --reload --port 8000Visit http://localhost:8000 for the web UI, http://localhost:8000/docs for API docs.
# Search packages
curl 'https://assay.tools/v1/packages?q=email&limit=5'
# Get a specific package
curl 'https://assay.tools/v1/packages/resend'
# Agent-optimized guide
curl 'https://assay.tools/v1/packages/resend/agent-guide'
# Compare packages
curl 'https://assay.tools/v1/compare?ids=resend,sendgrid,postmark'
# Stats
curl 'https://assay.tools/v1/stats'Assay includes an MCP server so AI agents can query ratings at runtime:
{
"mcpServers": {
"assay": {
"command": "python",
"args": ["-m", "assay.mcp_server"]
}
}
}Tools: find_packages, get_package, compare_packages, list_categories
Each package is evaluated across 5 weighted dimensions:
| Dimension | Weight | What it measures |
|---|---|---|
| MCP Quality | 20% | Tool descriptions, schema completeness, error handling |
| Documentation | 20% | Accuracy, examples, completeness for agent consumption |
| Error Messages | 15% | Actionable errors that help agents self-correct |
| Security | 15% | Auth patterns, input validation, least-privilege design |
| Auth Complexity | 15% | How easy it is for an agent to authenticate |
The AF Score is the weighted average, scaled 0-100.
Assay is an agentic business — AI agents handle evaluation, report generation, prospecting, and operational execution. See BUSINESS.md for the full business model, pricing tiers, product streams, and operating philosophy.
MIT