0130 points
Safety
Can an agent verify identity, permissions, and intent before it takes a destructive or irreversible action?
Unconfirmed writes · permission preflights
MCP agent-readiness scorecard
Your tools work.
Can an agent tell?
Gradable scans your MCP server and tells you, with a real number, whether an AI agent can actually use it correctly — or is silently guessing wrong.
- Real generated tasks
- Independent judge CLIs
- Actionable fixes
“The function returned 200” is not the same as “the agent understood the function.”
Gradable measures the interface your model actually sees.
The score
Four signals.
One honest number.
Gradable examines the tool surface, runs generated tasks through real agent workflows, and turns the evidence into a weighted score out of 100. It measures agent usability—not popularity, uptime, or stars.
0230 points
Legibility
Do names, descriptions, parameters, and return values tell the model exactly what each tool does?
Description defects · tool confusability
0320 points
Economics
How much context does the server consume before an agent has done any useful work—and is that cost justified?
Definition tokens · context overhead
0420 points
Discoverability
Given a real task, can an agent find and select the correct tool without guessing among near-duplicates?
Selection accuracy · task coverage
How it works
Static schemas meet dynamic judgment.
A useful score cannot come from counting fields alone. Gradable combines structural analysis with generated tasks and CLI-based judges, then identifies the highest-impact change you can make.
- 01Ingest the server
Capture the tools and the interface an agent receives.
- 02Generate real tasks
Test selection and behavior against representative intent.
- 03Judge the outcomes
Score safety, clarity, cost, and discoverability.
- 04Prioritize the fix
Turn the evidence into a scorecard your team can use.
Public leaderboard
The score is the start
of the conversation.
7 completed assessments
Latest scan Aug 1, 2026
| Rank | Server | Score | Grade | Safety | Legibility | Economics | Discovery | Scanned |
|---|---|---|---|---|---|---|---|---|
| 01 | Context7 (Upstash, live library docs lookup) context7-mcp · v3.2.5 | 90.6/100 | A | 30.0/30 | 30.0/30 | 17.8/20 | 12.8/20 | |
| 02 | Sequential Thinking (official MCP reference server) mcp-sequential-thinking · v0.2.0 | 67.4/100 | D | 0.0/30 | 30.0/30 | 17.4/20 | 20.0/20 | |
| 03 | Everything (official MCP protocol-feature demo server) mcp-everything · v2.0.0 | 64.5/100 | D | 0.0/30 | 27.7/30 | 17.4/20 | 19.4/20 | |
| 04 | Memory (official MCP reference server) mcp-memory · v0.6.3 | 56.6/100 | F | 0.0/30 | 24.5/30 | 16.4/20 | 15.7/20 | |
| 05 | Filesystem (official MCP reference server) mcp-filesystem · v0.2.0 | 50.1/100 | F | 0.0/30 | 25.5/30 | 16.1/20 | 8.6/20 | |
| 06 | Relaycast (headless Slack for AI agents) relaycast · v0.1.2 | 46.8/100 | F | 0.0/30 | 19.9/30 | 10.1/20 | 16.9/20 | |
| 07 | Agent Relay (this workspace's own MCP fleet server) agent-relay · v11.2.0 | 46.6/100 | F | 0.0/30 | 20.3/30 | 14.9/20 | 11.4/20 |
Category values are earned points: Safety and Legibility are scored out of 30; Economics and Discoverability out of 20. These reports use the same judge and prompt-template version, so their scores are directly comparable.
Initial assessment
Put your MCP
on the board.
Submit a public server URL or npm package. A full Gradable run executes generated tasks against the server and evaluates the outcomes with independent judge CLIs, so assessments run in a controlled environment rather than in your browser.
We’ll run the assessment and follow up with the score. This is the real pipeline—the same one behind the public leaderboard.
Submission received
You’re in the assessment queue.
Thanks—we’ll run the full assessment and follow up with your score.