About AgentRigs

AgentRigs is a public index of real AI-agent setups — the configuration people commit for Claude Code, Codex, Cursor, Gemini CLI and Copilot. Think dotfyle, but for agent rigs. The goal: make it easy to see what actually works in the wild, and copy it.

The current index has 6,413 rigs from 6,413 public repositories (76% carry an open-source license), last refreshed 2026-10-09.

Where the data comes from

• Repositories are discovered through GitHub's public search and topic pages (e.g. claude-code, dotfiles, claude-code-config, agent-skills) and curated "awesome" lists, plus submissions.
• For each repo we read the file tree and fetch only agent-config files: .claude/settings.json, .mcp.json, .cursor/mcp.json, .codex/config.toml, CLAUDE.md, AGENTS.md, GEMINI.md, .cursorrules, .cursor/rules/*, skills/*/SKILL.md frontmatter, .claude/agents/*.md frontmatter, command names and hook definitions.
• Forks and archived repos are skipped. Repo metadata (stars, license, last push) comes from GitHub's public API and ecosyste.ms.
• We store structure, not content. Instruction files are reduced to line and token counts. MCP env and header values are never stored — only key names. Every stored string passes a secret scanner (API-key, token, private-key, JWT, connection-string and high-entropy patterns); matches are replaced with <redacted>. Home-directory paths are shortened to ~.

Scoring method

Context tax

An estimate of the tokens a rig loads into every request before you type anything:

• Instructions: root-level CLAUDE.md / AGENTS.md / GEMINI.md / .cursorrules / always-apply Cursor rules, plus one level of @imports. We take the heaviest harness, not the sum. Tokens ≈ characters ÷ 4. Nested CLAUDE.md files in subfolders load lazily and are not counted.
• MCP tool schemas: a per-server estimate from typical tool counts (e.g. Context7 ≈ 1.2k, Playwright ≈ 7.5k, GitHub ≈ 18k tokens); unknown servers count as 2.5k. Harnesses with deferred tool loading (tool search) pay less — treat this as an upper bound.
• Skills & subagents: only name + description are preloaded, so we count (name + description) ÷ 4 + 8 tokens each.

It's a comparison tool, not a bill: real costs depend on caching, model and harness.

Guardrail score (0–5)

One point each for committed config that:

1. Blocks destructive commands — a deny/ask rule matching rm -rf, force-push, reset --hard, sudo, git clean, DROP, terraform destroy… or a PreToolUse hook that screens Bash.
2. Protects secrets — a deny rule for .env, keys, ~/.ssh, .aws, credentials…
3. Has a pre-tool screening hook — any PreToolUse hook.
4. Avoids YOLO mode — committed permission config without bypassPermissions, blanket Bash(*) allows, or Codex danger-full-access.
5. Sandbox or ask-first — sandbox enabled, Codex workspace-write/read-only, or explicit ask rules.

Rigs without committed permission config score low by design: harness defaults may still prompt you, but nothing in the repo enforces it.

Archetype

First match wins: YOLO Cowboy (explicit bypass) → Fort Knox (guardrails ≥4) → MCP Hoarder (6+ servers) → Orchestrator (5+ subagents) → Skill Collector (10+ skills) → Automator (4+ hooks) → Minimalist (<2.5k tokens, ≤4 components) → Pragmatist.

Component matching

MCP servers are identified by npm/PyPI package, Docker image, git source or remote host, with aliases merged (e.g. @upstash/context7-mcp and mcp.context7.com are both Context7). Local/custom servers are shown on rig pages but not ranked. Skills and subagents are matched by normalised name and listed once they appear in 4+ rigs (plugins 2+); MCP servers are listed from their first rig.

Rig Doctor, Trending, Starters and The Rig Report

Rig Doctor applies exactly this method to your own setup: paste CLAUDE.md, AGENTS.md, settings.json and .mcp.json (analysed in your browser, never uploaded), or give a public GitHub URL (a Pages Function fetches only the standard agent-config paths from raw.githubusercontent.com with size caps and timeouts, redacts env values and anything secret-looking, and stores nothing). You get your context tax by source, guardrail score, archetype, percentiles vs all indexed rigs, fixes with estimated token savings, ready-to-paste guardrail snippets and a shareable result card.

Trending compares adoption among rigs pushed in the last 7 / 30 days with older rigs, and keeps a dated snapshot of every component's adoption each day (sparklines and week-over-week change appear as snapshots accumulate). Starter rigs are curated, guarded setups distilled from well-scored public rigs, each citing its sources. The Rig Report is a free weekly email built from the same data; it has one clearly labelled sponsor slot per issue.

Badges

Every ranked component has a live SVG badge at https://agentrigs.dev/badge/<type>/<id>.svg, and every rig at https://agentrigs.dev/badge/rig/<owner>/<repo>.svg. Copy the snippet from any component or rig page.

Contact

Corrections, takedowns and partnerships: hello@agentrigs.dev · opt-out form. AgentRigs is independent and not affiliated with Anthropic, OpenAI, Anysphere (Cursor), Google or GitHub. Product names are trademarks of their owners.