Agent Catalog
Agent Passports document what each AI agent CAN do: what it reads, writes, accesses, spends, exposes, and what actions it gates behind human approval. These passports pair with package Capability Passports to compute compatibility at install time.
All v0.1 entries sourced from publicly documented agent behavior as of May 2026. Unverified until Phase 6 signing ships. Read the spec.
Claude Code
localby Anthropic
Agentic coding tool with direct shell, filesystem, and git access in the user's local environment. Supports configurable approval modes.
Codex CLI
localby OpenAI
OpenAI Codex CLI: local coding agent with shell access. Three approval modes (suggest, auto-edit, full-auto) control how much the agent acts without prompting.
Cursor
localby Anysphere
AI-enhanced code editor with inline agentic editing, multi-file context, terminal access, and Composer agent mode.
Gemini CLI
localby Google
Google Gemini CLI: open-source command-line AI agent with file, shell, and web access. Uses the Gemini API with a generous free tier.
Aider
localby Paul Gauthier (open source)
Open-source AI pair-programming CLI. Works with any git repository; commits changes automatically. Supports all major model APIs.
ChatGPT (browser/app)
sandboxedby OpenAI
ChatGPT in web browser or mobile app. Runs in OpenAI's hosted environment. File writes and code execution are confined to a session sandbox the user does not own.
Claude.ai (browser/app)
sandboxedby Anthropic
Claude.ai in web browser or mobile app. Runs in Anthropic's hosted environment. Code execution and file operations are session-sandboxed.
Gemini (web)
sandboxedby Google
Gemini at gemini.google.com. Runs in Google's hosted environment. Code execution is session-sandboxed; Workspace integrations require explicit user opt-in.
Grok (web)
sandboxedby xAI
Grok at grok.com or x.com. Runs in xAI's hosted environment. Session-sandboxed. Remote MCP connector support available to eligible accounts.
About these passports
- ·Agent Passports describe publicly documented behavior. APAI does not test agent runtime behavior and makes no certification claims.
- ·All entries carry
verified: falseuntil Phase 6 cryptographic signing ships. - ·v0.1 compatibility matching is descriptive. Runtime enforcement via Policy Pack lands in Phase 4.
- ·Agent Passport Spec v0.1 documents the full schema and field semantics.