For AI agents: a documentation index is available at /llms.txt — markdown versions of all pages are available by appending index.md to any URL path.

Tools

Every tool here exists to answer a research question, and every one is open source. The audit tools measure compliance with our specifications; the instrumentation toolchain gathers the empirical evidence our research is built on.

afdocs

afdocs.dev | v0.18.7

An open-source npm package that audits any documentation site against the Agent-Friendly Documentation Spec. Includes:

  • CLI tool for quick command-line audits (npx afdocs check https://docs.example.com)
  • Programmatic API for integration into CI/CD pipelines
  • Test helpers (vitest) for incorporating agent-friendliness checks into existing test suites
  • Multiple output formats: text, scorecard (category breakdowns, system-level diagnostics, and per-check fix suggestions), JSON, markdown, and GitHub Actions annotations

Implements all 23 checks of the spec across its 7 categories, with active development tracking the draft spec. Powers Fern’s Agent Score directory.

skill-validator

github.com/agent-ecosystem/skill-validator | v1.5.6

A stable, production-ready tool for analyzing Agent Skills at scale. Available via Homebrew (brew install agent-ecosystem/tap/skill-validator), Go install, or as a library for custom tooling. Features:

  • Structure validation: Spec compliance, token counts, orphan detection, internal link verification
  • Content analysis: Word count, code ratio, imperative ratio, specificity, and density metrics
  • Cross-contamination detection: Identifies programming language confusion across skill files
  • LLM-as-judge scoring: Six evaluation dimensions (clarity, actionability, token efficiency, scope discipline, directive precision, novelty) informed by research identifying novelty as a key predictor of skill value
  • Output formats: Text, JSON, markdown, and GitHub Actions annotations

The data behind the Agent Skill Report.

agentsummons

agentsummons.dev | v0.3.1

A Go library and CLI for invoking agent harnesses (Antigravity CLI, Claude Code, Codex CLI) in headless mode. Ships as a static Go binary, with npm and PyPI wrapper packages. The foundation for reproducible, scriptable agent testing:

  • Unified interface: One API across harnesses, abstracting away harness-specific flags and quirks
  • Multi-turn sessions: Session references in JSON output enable conversation resumption across separate invocations
  • Versioned JSON envelope: Standardized output schema for cross-language compatibility
  • Explicit over silent: Setting a field a harness can’t express raises an UnsupportedError instead of dropping the field
  • Drift detection: A doctor command validates installed harness versions against expected flag surfaces
  • Container-ready: Builds static binaries suitable for containerized environments

agentminutes

agentminutes.dev | v0.3.1

Meeting minutes for your agents: parses native agent harness session logs (Antigravity CLI, Claude Code, Codex CLI) into one unified, comparable event schema. Ships as a static Go binary, with npm and PyPI wrapper packages. The companion to agentsummons, turning raw transcripts into analyzable data:

  • Mechanical line-accounting: Every source line becomes an event, a skip, or an error, so every line is accounted for
  • Transcript discovery: Locates sessions by working directory, session ID, or time window
  • Standard vocabulary: Event schema aligned with Agent Client Protocol and OpenTelemetry conventions
  • Streaming adapters: Processes large transcript files incrementally
  • Precise errors: Failures identify the harness, version, and problematic lines

skillxp

skillxp.dev | v0.1.1

A skill invocation runner: stages an Agent Skill in a fresh, isolated fixture, invokes an agent harness headlessly, and reports what actually reached the model, with transcript evidence. Ships as a static Go binary, with npm and PyPI wrapper packages. Built on agentsummons and agentminutes:

  • Evidence, not self-reporting: Every claim traces to a session transcript, with line numbers
  • Provenance classification: Traced phrases are attributed to their source (harness-injected, tool results, model output)
  • Rates, not anecdotes: Repeated runs in fresh fixtures report loading behavior as rates across runs
  • Powers Agent Skill Implementation: The same toolchain drives that benchmark’s runner and its 38 automated checks across real harnesses

agent-traffic-classifier

github.com/agent-ecosystem/agent-traffic-classifier | v0.2.0

Classifies web traffic from server log entries and header capture into human, AI agent, AI crawler, AI assistant, programmatic, and other bot categories. No analytics service or middleware required. Gives site owners direct visibility into how much of their traffic comes from agents, supporting the observability checks in the Agent-Friendly Documentation Spec.

Research Built in the Open

The collective is independent and self-funded by design. Nobody pays for our conclusions, and everything we publish (specifications, tools, scoring algorithms, and research methodology) is open source and freely available.

Follow the Work on GitHub