Works with
Interoperability
Connect the suite to the agent stacks and protocols you already use. Statuses are what we tested. No logos. No implied endorsement or affiliation.
How to read a row
- TESTED — we ran the path against the published open-source tool or config (usually with a mock model).
- AVAILABLE — a recipe, adapter, or file ships; launch or offline coverage only.
- DOCUMENTED — research notes; no adapter yet, or an honest gap.
Our runs are TESTED, never VERIFIED. Product use does not wait on external reviews.
Matrix
| Stack | Works with | How tested | Status |
|---|---|---|---|
| Model Context Protocol | gaig-mcp; Claude Code, Desktop, Cursor, VS Code, Codex, Gemini CLI, LangGraph | stdio ClientSession; tools list; authority_effect: NONE | TESTED |
| MCP registry | server.json on ai-governance-mcp | Validated against the published registry schema; launch args pin v0.5.1 | TESTED |
| OpenAI Agents SDK | openai_agents_guard.py guardrail + require_approval | Offline with openai-agents 0.24.0, real gaig-mcp stdio, scripted model; Level 5 denied, Level 4 paused; not against a live model | TESTED |
| OpenAI Codex CLI | ~/.codex/config.toml | One recorded local stdio run (Codex CLI 0.151.0) | TESTED |
| OpenAI AGENTS.md | Suite AGENTS.md files | Machine front door + suite path | AVAILABLE |
| Anthropic MCP / Claude Code | Spec-shaped server; PreToolUse hooks | Hook dry-run offline; launch tested | TESTED / AVAILABLE |
| Anthropic Claude Agent SDK | claude_agent_sdk_guard.py PreToolUse hook | Offline with claude-agent-sdk 0.2.174; hook called with recorded-format inputs; not run against Claude | TESTED |
| Anthropic SKILL.md | Skill files in MCP and Agentic | File present; not run inside Claude | AVAILABLE |
| Google A2A | A2A passport extension + agent card notes | Schema + docs; offline | AVAILABLE |
| Google ADK / Gemini CLI | Gemini CLI MCP recipe | Launch command; not inside Gemini | AVAILABLE |
| Microsoft Agent Framework / AutoGen / Semantic Kernel | MCP stdio client pattern | Documented; VS Code mcp.json launch tested | AVAILABLE |
| VS Code MCP | .vscode/mcp.json | Launch tested; not inside VS Code | AVAILABLE |
| LangGraph | langgraph_passport_gate.py | Offline with langgraph 1.2.17 and synthetic passports; no model | TESTED |
| CrewAI | crewai_guard.py before/after tool hooks | Offline with crewai 1.15.33; hook invoked with real ToolCallHookContext; not against a live model | TESTED |
| LlamaIndex / smolagents | Same Action Level guard pattern as CrewAI | Pattern only; no dedicated adapter yet | DOCUMENTED |
| AWS Strands / AgentCore | MCP client notes | Research only | DOCUMENTED |
| OpenTelemetry GenAI | Span attributes for check calls | Example exporter; offline | AVAILABLE |
| AAIF / AGNTCY | Alignment notes | Research only | DOCUMENTED |
| xAI | OpenAI-compatible Lead Agent serve; Grok CLI MCP recipe | Launch + mock; not inside Grok | AVAILABLE |
| SpaceX | — | No public agent SDK found | DOCUMENTED (gap) |
Lead Agent
The browser Lead Agent runs a deterministic mock model with no account or key. Optional local serve from Agentic AI Governance talks to OpenAI-, Anthropic-, or xAI-compatible endpoints when keys sit in the local environment. Keys never enter the browser.
Limitations
- Adapters wrap Action Levels and passports; they do not authorize any action.
- Client-inside sessions are mostly launch-tested, not session-tested.
- Evidence gates that wait on outside parties stay unchanged and are not required for these paths.