AgenticScore

Score any API's readiness for AI agent consumption.

AI agents (Claude, GPT, custom agents) are consuming APIs at scale — but most APIs weren't designed for machine callers. AgenticScore scores your OpenAPI spec from 0-100 across 6 dimensions that determine how well AI agents can use your API without human intervention.

Quick Start

npx agenticscore score your-api.yaml

Output:

  🟠 Agent Readiness: 47/100 (Grade D)
  My API v2.1.0
  23 operations, 15 schemas

  Categories:
    examples     ████░░░░░░ 40%
    semantics    █████░░░░░ 50%
    intent       ██████░░░░ 60%
    errors       ███░░░░░░░ 30%
    parameters   █████░░░░░ 55%
    pagination   ██░░░░░░░░ 20%

  Top Issues:
    • 18/23 operations lack examples
    • 12/23 operations don't document error responses
    • 3 list operations lack pagination parameters

Install

# Global install
npm install -g agenticscore

# Or run directly
npx agenticscore score api.yaml

Scoring Categories

Category	Weight	What It Measures
Examples	25%	Do operations include request/response examples for agents to learn from?
Semantics	33%	Are descriptions clear enough for agents to understand intent?
Intent	17%	Do operationIds clearly express what each endpoint does?
Errors	15%	Are failure modes documented so agents can handle them?
Parameters	8%	Are query/path parameters described?
Pagination	7%	Can agents iterate list endpoints?

CI/CD Integration

Fail your build if agent readiness drops below a threshold:

agenticscore score api.yaml --min-score 70

Exit code 1 if score is below the threshold.

GitHub Actions

- name: Check API Agent Readiness
  run: npx agenticscore score ./openapi.yaml --min-score 70

Output Formats

# Pretty terminal output (default)
agenticscore score api.yaml

# JSON (for programmatic use)
agenticscore score api.yaml --format json

# Markdown (for PRs/docs)
agenticscore score api.yaml --format markdown

Why This Matters

AI agents need APIs that are:

1. Self-documenting — Agents can't read your wiki. They need descriptions in the spec.
2. Example-rich — Agents learn request/response shapes from examples, not guesswork.
3. Predictable in failure — RFC 9457 Problem Detail lets agents parse and retry intelligently.
4. Navigable — Pagination, filtering, and sorting must be documented for agents to iterate data.
5. Intent-clear — listActiveUsers tells an agent what it does. get_1 does not.

The 5 Fixes That Move Any API from D to A

1. Add request/response examples to every operation
2. Write 2-sentence descriptions (what it does + when to use it)
3. Document error responses with RFC 9457 Problem Detail
4. Use descriptive operationIds (listUsers, not get_1)
5. Document pagination parameters on list endpoints

Programmatic Usage

import { scoreSpec } from "agenticscore";
import { readFileSync } from "fs";

const spec = readFileSync("api.yaml", "utf-8");
const report = scoreSpec(spec);

console.log(report.overallScore); // 0-100
console.log(report.grade);        // A, B, C, D, F
console.log(report.topFindings);  // Actionable fix list

Custom Rules (Coming Soon)

import { scoreSpec } from "agenticscore";
import { rules } from "agenticscore/rules";

// Add your own rules
const customRules = [
  ...rules,
  {
    id: "my-rule",
    name: "Custom Check",
    category: "semantics",
    weight: 5,
    description: "My organization's requirement",
    evaluate: (spec) => ({ score: 1, findings: [] }),
  },
];

License

MIT