IDE runtime parity
AgentKeeper supports several coding-agent and desktop-agent surfaces, but they do not all expose the same enforcement hooks. Treat Claude Code as the baseline, then use this matrix to decide...
AgentKeeper supports several coding-agent and desktop-agent surfaces, but they do not all expose the same enforcement hooks. Treat Claude Code as the baseline, then use this matrix to decide whether a control is a blocking control, an audit control, a managed-configuration control, a preview claim, or unsupported by hooks.
Last reviewed: June 3, 2026.
What parity means
Runtime parity means AgentKeeper can evaluate the action before it happens and return a vendor-supported blocking decision. It does not mean OS-level interception, and it does not mean a post-tool hook can undo work that already happened.
Source-backed matrix
| Surface | Current docs | Best-supported blocking path | Known limitation |
|---|---|---|---|
| Claude Code | Hooks reference | PreToolUse, UserPromptSubmit, Stop, and managed policy hooks | Baseline surface. |
| Codex | Hooks, managed configuration | PreToolUse, PermissionRequest, UserPromptSubmit, Stop, and managed requirements.toml hooks | Post-tool feedback cannot undo completed commands. |
| GitHub Copilot | Hooks reference | preToolUse for CLI/cloud, CLI-only permissionRequest, and agent stop continuation | VS Code and extension behavior must be version validated before customer rollout. |
| Cursor | Hooks docs | preToolUse, shell/MCP before hooks, prompt hook, stop hook, MDM or Enterprise cloud hook distribution | CLI coverage is version-specific. |
| Windsurf | Cascade hooks | pre_run_command, pre_write_code, pre_read_code, pre_mcp_tool_use, and pre_user_prompt | Only pre-hooks block; warn-only UI is audit-only. |
| Gemini CLI | Configuration, writing hooks | BeforeTool, BeforeAgent, AfterAgent, SessionStart, and structured deny/block JSON | Fleet rollout requires managed settings/status hardening. |
| Google Antigravity | AgentKeeper setup guide | PreToolUse, PostToolUse, PreInvocation, PostInvocation, Stop, and structured allow/deny/approval-request responses | IDE-only coverage. No browser-extension capture or OS-level interception is claimed. |
| Claude Chat | Anthropic MCP overview, MCP user quickstart, AgentKeeper setup guide | MCP Gateway routing from Claude Desktop claude_desktop_config.json | Covers routed MCP tools. It is not a Claude-Code-style hook surface. |
| Claude Cowork | Getting started, plugins | AgentKeeper plugin guardrail for emitted Cowork events | Official Cowork docs do not document Claude-Code-style hook parity. |
Control labels
The dashboard uses these labels for IDE badges:
| Label | Meaning |
|---|---|
| Block | The vendor exposes a pre-action hook or lifecycle decision AgentKeeper can use to stop the action. |
| Managed config | The vendor supports admin-enforced config for the control, usually stronger than local user config. |
| Preview | The vendor documents the concept, but rollout must be validated by product version, surface, or customer tenant. |
| Audit | AgentKeeper can record the event or post feedback, but should not promise user-visible blocking. |
| Unsupported by hooks | The vendor docs do not expose a hook-level control for that action. |
Rollout guidance
Start new enterprise rollouts with Claude Code or Codex when the customer requires the closest documented parity with managed hooks. Use Gemini CLI, Google Antigravity, or Windsurf when shell and file enforcement are the priority and local CLI/IDE install is acceptable. Use Antigravity for Antigravity IDE command, file, URL, MCP, subagent, schedule, and invocation-lifecycle coverage only; do not describe it as browser-extension capture. Use Cursor where its app hooks are available and version validated. Use Copilot for CLI/cloud-agent controls first, then validate VS Code behavior before claiming full local IDE parity.
Claude Chat remains valuable for Claude Desktop MCP governance through Gateway. Cowork remains valuable for Claude Desktop Local Agent Mode governance, but it should be described as Cowork plugin coverage rather than full coding-IDE parity.
Shadow AI managed preseed detection
For managed macOS and Windows fleets, Jamf, Iru/Kandji, and Intune can deploy AgentKeeper-owned hook/config coverage before an IDE runs. The Native Linux RPM provides the same discovery and reconciliation model for explicit enrolled users across the nine-surface catalog. Managed reports reach POST /api/v1/runtime/managed-coverage, which lets AI Discovery distinguish mdm_preseed and native_linux_rpm provenance when a known AI IDE is present or an expected hook is missing.
This signal is setup and inventory evidence. It does not uninstall apps, block launches, or change runtime policy evaluation. A pure preseeded row without app presence is coverage evidence; a row with app presence becomes a reviewable Shadow AI candidate unless the surface is approved or accepted risk.
Copilot remains workspace-limited: AgentKeeper can report user/workspace hook coverage, but it does not mutate arbitrary repositories to create workspace hooks.
Fleet deployment
On supported RHEL and Rocky Linux hosts, the Linux Universal RPM is the primary fleet path for all nine catalog surfaces. It reports absent tools as not_seen, preserves unsupported combinations such as Cowork on RHEL/Rocky as unsupported or discovery_only, and does not upgrade any surface to active until a real runtime event arrives.
On macOS, deploy the universal signed PKG through Jamf, Iru/Kandji, or Intune. One package reconciles AgentKeeper-owned entries for Claude Code, Cursor, Windsurf, Copilot, Codex, Gemini CLI, and Antigravity; do not deploy per-IDE MDM installers beside it.
Claude Chat uses MCP Gateway, and Cowork uses the
Cowork plugin. Those remain separate from the universal
endpoint package. AI Discovery distinguishes preseeded configuration from a
real active event, so neither package success nor preseeded alone proves live
protection.