Keyboard shortcuts

Press or to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

SaaS Coding-Agent Governance Limits

SaaS coding agents (Claude.ai, ChatGPT, Cursor cloud) run in opaque cloud environments. This imposes hard limits on what Agent Assembly can govern: in-process enforcement (L2/L3) is structurally impossible at the SaaS boundary. All SaaS adapters are capped at L1Observe.

The table below documents each capability per provider so operators can set accurate expectations and plan compensating controls where needed.

ProviderCapabilityStatusReason
Claude.aiMCP allowlist✅ SupportedWorkspaces API exposes MCP configuration
Claude.aiSystem-prompt overlay⚠️ PartialCan prepend governance note; operator must apply manually
Claude.aiNetwork egress deny❌ UnsupportedSaaS boundary; no network-layer hook
Claude.aiL2 enforcement❌ UnsupportedSaaS boundary prevents in-process enforcement
ChatGPTMCP allowlist❌ UnsupportedEnterprise API does not expose MCP configuration
ChatGPTSystem-prompt overlay⚠️ PartialCustom GPT system-prompt field; operator applies
ChatGPTNetwork egress deny❌ UnsupportedSaaS boundary; no network-layer hook
ChatGPTL2 enforcement❌ UnsupportedSaaS boundary prevents in-process enforcement
Cursor cloudMCP allowlist❌ UnsupportedAudit webhook does not expose MCP config
Cursor cloudSystem-prompt overlay❌ UnsupportedNo system-prompt surface in audit webhook
Cursor cloudAudit event ingestion✅ SupportedSigned audit webhook delivers all agent actions
Cursor cloudL2 enforcement❌ UnsupportedSaaS boundary prevents in-process enforcement

Compensating controls

For capabilities marked ❌ Unsupported, operators should consider:

  • Network egress deny: use the sidecar proxy (aa-proxy) or eBPF layer on the host machine that runs the SaaS agent’s browser or desktop client.
  • L2 enforcement: not possible at the SaaS boundary. Governance relies on webhook audit events and operator-applied configuration overlays.
  • MCP allowlist (ChatGPT, Cursor): use network-layer controls to restrict which MCP servers the agent host can reach.

References

  • Limitations and Known Bypasses — the equivalent honest-boundaries discussion for the locally-running CLI/extension adapters.
  • L0-L3 Capability Matrix — the maintained matrix for the CLI/extension dev-tool adapters (Codex, GitHub Copilot, Windsurf Cascade, Claude Code). Its own SaaS Coding-Agent row is a separate placeholder pending AAASM-918; this page is the current source for what the shipped aa-devtool-saas overlays actually do.
  • Legacy per-tool capability detail (superseded by the matrix above): Claude Code, Codex CLI, GitHub Copilot, Windsurf Cascade.
  • aa-devtool-saas crate — overlay::ClaudeAiOverlay, overlay::ChatGptOverlay, and the per-provider webhook signature verification this page’s Cursor-cloud row describes.

Last updated: 2026-08-01 by Chisanan232