AI Capability Reference

Previously known as "AI Feature Tracker." Renamed to "AI Capability Reference" in February 2026 to better reflect the project's ontology-first, capability-centric approach.

Plain-English reference for AI capabilities, plans, constraints, and implementations.

---

What is this?

A single source of truth for answering questions like:

• "Is ChatGPT Agent Mode available on the $8/mo plan?" (No, Plus or higher)
• "Can I use Claude Cowork on Windows?" (Not yet, macOS only)
• "Which open/self-hosted models can I realistically run on my hardware?" (Depends on VRAM)

Built for fellow AI facilitators, educators, designers, and anyone who needs accurate, current information about AI tool availability.

How it's built

There is no database behind this project. Every piece of data — features, pricing tiers, platform support, talking points — lives in plain markdown files under data/. A single build script (node scripts/build.js) reads those files and renders the entire static site into docs/.

That's the whole stack: markdown files, javascript, and Git.

This means contributing doesn't require a dev environment, an ORM, or a running database. You edit a .md file, open a PR, and the CI rebuilds the site. Git provides versioning, review, and an audit trail for every change. If you can read a markdown table, you can read (and fix) the data.

Scope

This reference covers:

• major consumer-facing AI products with meaningful public usage or visibility
• commercially available AI systems that ordinary people can sign up for and use
• important self-hosted/open model families and runtimes that people can realistically run or choose locally

This reference does not aim to catalog every enterprise AI vendor, infrastructure platform, or niche model release.

The ~1% market share idea is used here as a practical inclusion heuristic, not a strict ontology field, because public usage data is inconsistent and is usually measured at the product level rather than the model level.

All prices are listed in USD. Feature availability and pricing reflect the United States region by default; availability may differ in other regions.

Platforms Covered

Platform	Vendor	Features Tracked
ChatGPT	OpenAI	Agent Mode, Canvas, Voice, Sora, DALL-E, Search, Deep Research, Codex, Custom GPTs
Claude	Anthropic	Code, Cowork, MCP, Connectors, Projects, Artifacts, Extended Thinking, Vision
Copilot	Microsoft	Office Integration, Designer, Vision, Voice, Agent Builder
Gemini	Google	Advanced, NotebookLM, AI Studio, Deep Research, Gems, Workspace, Imagen, Veo, Live, Project Astra
Perplexity	Perplexity AI	Comet Browser, Agent Mode, Pro Search, Focus, Collections, Voice
Grok	xAI	Chat, Aurora (images), DeepSearch, Think Mode, Voice
Meta	Meta	Llama 3.3, Llama 4
Mistral	Mistral	Codestral, Mistral Large/Nemo, Mistral Small 3
DeepSeek	DeepSeek	DeepSeek V3 / R1
Alibaba	Alibaba	Qwen 2.5, Qwen 3, Qwen 3.5, Qwen-Coder
Ollama	Ollama	Self-hosted runtime product
LM Studio	LM Studio	Self-hosted runtime product

Features

• Plan-by-plan availability — See exactly which tier unlocks each feature
• Capability-first view — Browse plain-English capabilities in addition to the feature view
• Platform support — Windows, macOS, Linux, iOS, Android, web, terminal, API
• Talking points — Ready-to-use sentences for presentations (click to copy)
• Category filtering — Voice, Coding, Research, Agents, and more
• Price tier filtering — Find features at your budget
• Provider toggles — Focus on specific platforms
• Dark/light mode — Toggle for your preference
• Permalinks — Link directly to any feature with shareable URLs
• Shareable URLs — Filter state preserved in URL parameters
• Maintainer-led — Managed by SnapSynapse; public issues and PRs welcome

Accessibility

This site is designed to meet WCAG 2.1 AA standards:

• Keyboard navigation — Full keyboard support with ↑/↓/j/k to navigate cards, Enter to copy, Tab to move between interactive elements
• Skip link — "Skip to main content" link for screen reader users (visible on focus)
• Focus indicators — Clear 2px accent-colored outlines on all interactive elements
• Color contrast — Minimum 4.5:1 contrast ratio for all text in both light and dark modes
• Reduced motion — Animations and transitions disabled when prefers-reduced-motion is enabled
• Touch targets — Minimum 44px touch targets on mobile for easier tapping
• ARIA attributes — Live regions announce filter count changes, decorative images marked with aria-hidden
• Semantic HTML — Proper heading hierarchy, landmark regions, and button/link semantics

If you're interested in accessibility tooling, check out skill-a11y-audit — a companion project that automates WCAG audits as a reusable AI skill. If you like what you see here, you'll love that.

Update Cycle

Data in this reference is kept current through a semi-automated process combining scheduled AI verification with human review.

How it works:

Every Sunday, a four-model cascade queries Gemini, Perplexity, Grok, and Claude to cross-check all tracked features — pricing tiers, platform availability, status, gating, and regional restrictions. To prevent provider bias, models are skipped when verifying features from their own platform (e.g. Gemini is not asked about Google features). A change is only flagged when at least three models agree on a discrepancy. Confirmed changes are surfaced as GitHub issues or pull requests for human review — nothing is auto-merged.

Link integrity is checked separately every Saturday, catching broken or redirected source URLs across all platforms.

Features are also marked with a Checked date. Anything not re-verified within 7 days is treated as stale and prioritised in the next run.

How to Contribute

Found outdated info? Want to add a feature? See CONTRIBUTING.md.

Quick version:

1. Edit the relevant record in data/platforms/, data/model-access/, data/products/, or data/implementations/
2. Include or preserve the evidence source link
3. Run node scripts/sync-evidence.js
4. Run node scripts/validate-ontology.js
5. Submit a PR

Automated Verification

This project includes an automated feature verification system that uses multiple AI models to cross-reference all feature data.

What gets verified

• Pricing tiers — Which subscription plans have access
• Platform availability — Windows, macOS, Linux, iOS, Android, web, terminal, API
• Status — GA, Beta, Preview, Deprecated
• Gating — Free, Paid, Invite-only, Org-only
• Regional availability — Global vs region-restricted features
• URLs — Feature page links are valid and accessible

How it works

1. Multi-model cascade — Queries Gemini, Perplexity, Grok (X/Twitter), and Claude
2. Bias prevention — Skips same-provider models (e.g., won't ask Gemini about Google features)
3. Consensus required — Needs 3 models to confirm a change before flagging
4. Auto-changelog — Confirmed changes are logged to each feature's changelog
5. Human review — Creates issues/PRs for review, never auto-merges

Running verification

# Verify all features
node scripts/verify-features.js

# Verify a specific platform
node scripts/verify-features.js --platform claude

# Check only stale features (>7 days since last check)
node scripts/verify-features.js --stale-only

# Dry run (no issues created)
node scripts/verify-features.js --dry-run

Link checking

Two link checkers serve different purposes:

CI checker (check-links.js) — runs in GitHub Actions weekly. Uses canonical categories from the collaboration protocol:

• ok
• broken
• soft-blocked (e.g., persistent 403/bot protection; informational)
• rate-limited (429; informational)
• timeout
• needs-manual-review

The checker fails CI on actionable problems (broken, timeout) while keeping soft-blocked/rate-limited as non-actionable signals.

node scripts/check-links.js              # Check all links
node scripts/check-links.js --broken-only # Show only broken links

Browser checker (check-links-browser.js) — runs locally through a real Chrome browser via Chrome DevTools Protocol. Bypasses all bot protection, captures page titles for content verification, and shows redirects. Zero external dependencies.

# Terminal 1: Start Chrome with remote debugging
# (--user-data-dir avoids conflicts with your normal browser session)

# macOS (Chrome)
/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome \
  --remote-debugging-port=9222 --user-data-dir=/tmp/chrome-link-check

# macOS (Brave, Edge, or any Chromium browser also works)

# Linux
google-chrome --remote-debugging-port=9222 --user-data-dir=/tmp/chrome-link-check

# Terminal 2: Run the checker
node scripts/check-links-browser.js                # All platforms
node scripts/check-links-browser.js -p claude       # One platform
node scripts/check-links-browser.js --help          # All options

Requires API keys (verification only): GEMINI_API_KEY, PERPLEXITY_API_KEY, XAI_API_KEY, ANTHROPIC_API_KEY

See VERIFICATION.md for full documentation.

Secret Scanning

This repo includes a lightweight secret scan for the current tree and full Git history.

node scripts/scan-secrets.js

Use --current-only or --history-only to narrow the scan when needed.

Local Development

# Clone the repo
git clone https://github.com/snapsynapse/ai-capability-reference.git
cd ai-capability-reference

# Build the dashboard
node scripts/build.js

# Sync ontology-native evidence records
node scripts/sync-evidence.js

# Validate ontology records
node scripts/validate-ontology.js

# Open it
open docs/index.html

Archived platform bundles live under data/archive/platforms/ for historical reference, and node scripts/build.js reads only active evidence from data/platforms/. Ontology-native evidence records live in data/evidence/index.json and are seeded with node scripts/sync-evidence.js.

Skills

Canonical cross-platform skill sources live under skills.

All repo skills should use skills/<name>/src/ as the only canonical source location. Do not commit mirrored skill copies under .claude/skills/, .perplexity/skills/, or other platform-local folders; those are local install locations, not source of truth.

Repo skills also assume the companion skill-provenance metaskill is available anywhere the bundles are imported. Treat it as a required dependency for bundle identity, manifest continuity, and cross-surface handoff.

Build platform exports with:

node scripts/build-skill-bundles.js

That generates Perplexity, Claude, and Codex-ready outputs under each skill's local dist/ directory without turning the repo itself into a pile of hand-maintained exports.

Data Format

Platform data is stored in simple markdown files. Example:

## Feature Name

| Property | Value |
|----------|-------|
| Category | agent |
| Status | ga |

### Availability

| Plan | Available | Limits | Notes |
|------|-----------|--------|-------|
| Free | ❌ | — | Not available |
| Plus | ✅ | 40/month | Message limit |

### Talking Point

> "Your presenter-ready sentence with **key details bolded**."

### Sources

- [Official docs](https://example.com)

See data/_schema.md for the full specification.

Deployment

The site auto-deploys via GitHub Actions when changes are pushed to main.

How it works

1. Build job (.github/workflows/build.yml)

   • Runs node scripts/build.js to regenerate all pages under docs/
   • If output changed, commits it back to main with [skip ci] to prevent loops
   • Runs on both pushes and PRs (PRs only validate the build, no commit)

2. Deploy job (same workflow)

   • Uploads docs/ folder to GitHub Pages
   • Only runs on pushes to main, not PRs

3. FTP deploy (.github/workflows/deploy-ftp.yml)

   • Parallel deployment to snapsynapse.com via locked ftp
   • Requires FTP_HOST, FTP_USER, FTP_PASS secrets

GitHub Pages setup

To enable GitHub Pages on a fork:

1. Go to Settings → Pages
2. Under "Build and deployment", select GitHub Actions
3. The workflow will deploy to https://<username>.github.io/ai-capability-reference/

Manual build

node scripts/build.js

Output files:

• docs/index.html — Capability homepage
• docs/implementations.html — Detailed Availability (implementation matrix)
• docs/constraints.html — Access & Limits (constraint view)
• docs/capabilities.html — Redirect stub (backward compat)
• docs/about.html — About page (generated from README.md)

License

MIT - see LICENSE

Credits

Created by SnapSynapse for the AI training community. With help from Claude Code, of course.

---

Found an error? Open an issue or submit a PR!