Capability matrix

Per-connector breakdown of block capability, native ask events, fail-closed support, subprocess policy, OpenShell sandbox tier, and HITL behaviour. The single source of truth for "can this connector do X?"

This page answers "can this connector do X?" for every connector. The table is checked against the connector code on every change. For what works on your operating system, see the support matrix.

Yes the connector supports itNo it doesn'tScroll the table sideways on a narrow screen.

ConnectorFamilyTool inspectionSubprocess policyBlockNative askFail-closedOpenShell sandbox
Claude Code
claudecode
hookspre-execution + response-scannoneYesYes
PreToolUse
Yesmanaged tier
verified end to end
HITL: Claude Code supports native PreToolUse ask prompts. CRITICAL findings still block; HIGH findings can pause for approval.
Codex
codex
hookspre-execution + response-scannoneYesNoYesmanaged tier
verified end to end
HITL: Codex has no native ask surface here; confirm becomes an alert/systemMessage with raw_action preserved. The TUI can review the event but cannot resume it.
OpenClaw
openclaw
proxypre-execution + response-scanshimsYesYes
before_tool_call
Yespending
HITL: OpenClaw supports DefenseClaw approval prompts for tool actions. Approvals reach chat-origin sessions via the bundled plugin.
Cursor
cursor
hookspre-execution + response-scannoneYesNoYesmanaged tier
unverified: cannot run yet
HITL: DefenseClaw does not enable Cursor native human approval. Confirm verdicts remain attributed alerts and cannot resume the original hook call.
Hermes
hermes
hookspre-execution + response-scannoneYesNoYes
Timeout or spawn failure can allow
user tier
verified end to end
HITL: Hermes has no native human-approval surface; confirm verdicts are recorded and alerted but cannot pause or resume the hook.
OpenCode
opencode
hookspre-execution + response-scannoneYesNoYesuser tier
unverified: cannot run yet
HITL: No native human-approval surface; blocks by throwing in the bridge plugin's tool.execute.before. confirm verdicts fall back to allow.
Amp
amp
hookspre-execution + response-scannoneYesYes
tool.call, tool.result
Yesuser tier
unverified: cannot run yet
HITL: Amp can ask in the active foreground thread during tool.call and before a tool.result reaches the model. Denial, a background thread, or unavailable plugin UI rejects the call or withholds the result rather than silently allowing it.
GitHub Copilot CLI
copilot
hookspre-execution + response-scannoneYesNoYes
Timeout or spawn failure can allow
managed tier
verified end to end
HITL: Copilot CLI confirmations block because its hook does not prove an approval prompt is available; VS Code Local uses its separate native ask.
Antigravity
antigravity
hookspre-executionnoneYesYes
PreToolUse
Yes
Timeout or spawn failure can allow
user tier
verified end to end
HITL: Antigravity documents native ask on PreToolUse only. DefenseClaw sends force_ask, which ignores cached Always Allow grants, never plain ask. DefenseClaw does not claim that this response overrides permission-bypass flags without persisted official-client evidence.
Devin
devin
hookspre-execution + response-scannoneYesNoYesuser tier
unverified: cannot run yet
HITL: Can block documented events but has no native resumable human-approval surface; confirm verdicts fall back explicitly.
Kiro
kiro
hookspre-execution + response-scannoneYesNoYesuser tier
verified end to end
HITL: Kiro has no native ask verdict; policy confirms are recorded and downgraded to alerts. ACP permission requests stay native client interactions and may be denied before the client displays them.
ZeptoClaw
zeptoclaw
proxypre-execution + response-scanshimsYesNoYespending
HITL: ZeptoClaw has no native ask surface; confirm uses its explicit fallback with raw_action preserved for TUI/audit review. There is no resumable approval.

Reading the matrix

ColumnMeaning
Familyproxy: DefenseClaw sits in the model data path. hooks: DefenseClaw uses the agent's hooks or plugin and the agent talks to its upstream directly.
Tool inspectionWhen DefenseClaw sees a tool call. pre-execution + response-scan scores it before it fires and again when it returns.
Subprocess policyshims: setup installs PATH shims for curl, wget, ssh, nc, pip and npm that ask the gateway before running the real binary. They do not isolate the process. none: the agent runs commands itself.
BlockWhether the connector's hooks can return a block decision at all.
Native askWhether the hooks can prompt you inside the agent. Without it, a confirm verdict becomes an alert, allow or context message that the TUI can show but cannot resume.
Fail-closedWhether DefenseClaw can deny a delivery failure in mode closed. A note marks agents that can still allow a call if the hook cannot start or times out.
OpenShell sandboxTamper tier and verification of the harness image. See below.
HITLHow human approvals reach you on that connector.

Two rules sit on top of the columns.

Antigravity asks and blocks only before a tool runs

Antigravity asks and blocks only on PreToolUse. PreInvocation and PostInvocation only change context, PostToolUse only observes, and Stop is always allowed.

A prompt rule blocks only where the agent can block a prompt

Claude Code, Codex, Cursor and Devin can. Kiro blocks a prompt only in Kiro IDE. Copilot CLI, Hermes, OpenCode, Amp and Antigravity have no such hook, so a matching prompt reaches the model. Tool-call rules still block there.

Common patterns

"I want maximum safety on a single connector"

Pick one with Block, Native ask and Fail-closed all yes, and a managed-tier sandbox verified end to end. Of the current connectors that is Claude Code: it asks inside the agent, and in a sandbox its hooks sit in a root-owned policy file.

"I want my agent's own hooks to enforce"

Claude Code, Codex, Cursor, Hermes, Devin, Copilot CLI, Antigravity, OpenCode and Kiro enforce through the agent's own hooks. Amp uses a policy plugin. None of them can block a request the agent has not yet shown to its hook or plugin surface.

"I want HITL approvals to surface inside the agent UI"

Claude Code (PreToolUse, in Manual mode only; other permission modes block), Antigravity (PreToolUse only), Amp (tool.call and model-bound tool.result in the active foreground thread) and OpenClaw can ask inside the agent. Copilot CLI is not counted: its hook does not prove that an approval prompt is available, so DefenseClaw blocks action-mode confirmations. The separate VS Code Local harness retains native ask. Cursor is not counted: DefenseClaw does not turn on its native ask. Every other confirm verdict takes a connector-specific fallback, and there is no TUI approval queue.

"I want to run the agent in an OpenShell sandbox"

defenseclaw sandbox run <harness> runs a coding agent in an NVIDIA OpenShell sandbox that sees only your project folder, on Linux or in a MicroVM on an Apple-silicon Mac (how). DefenseClaw builds images for the harnesses below. Every other connector is pending.

ConnectorImage pinTamper tierHook config in the imageVerification
Claude Code
claudecode
2.1.156managed/etc/claude-code/managed-settings.d/50-defenseclaw.jsonVerified end to end
Codex
codex
0.146.0managed/etc/codex/requirements.tomlVerified end to end
Cursor
cursor
2026.07.23-e383d2bmanaged/etc/cursor/hooks.jsonUnverified: cannot run yet. The Cursor Agent CLI needs a Cursor account (CURSOR_API_KEY or cursor-agent login) before any agent turn.
Not tested live: defenseclaw-cursor, in-sandbox login
Hermes
hermes
0.19.0user/etc/hermes/config.yamlVerified end to end
Not tested live: defenseclaw-openai, defenseclaw-anthropic
OpenCode
opencode
2.1.4user/etc/opencode/opencode.jsonUnverified: cannot run yet. V2-only image and bridge; previous V1 sandbox certification does not apply. A fresh V2 OpenShell ingress, block, and egress verification is required..
Amp
amp
0.0.1785334225-g9abe75user/sandbox/.config/amp/plugins/defenseclaw.tsUnverified: cannot run yet. Amp needs an Amp account API key (AMP_API_KEY) for every run.
Not tested live: defenseclaw-amp
GitHub Copilot CLI
copilot
1.0.88managed/etc/github-copilot/policy.d/50-defenseclaw.jsonVerified end to end
Not tested live: defenseclaw-copilot-github
Antigravity
antigravity
1.2.12user/sandbox/.gemini/config/hooks.jsonVerified end to end
Not tested live: defenseclaw-gemini
Devin
devin
3000.4.25user/sandbox/.config/devin/config.jsonUnverified: cannot run yet. The Devin CLI needs a Devin account login (devin auth login, a browser or pasted-token flow) before any agent turn.
Not tested live: in-sandbox login
Kiro
kiro
2.24.1user/usr/local/lib/defenseclaw/kiro/defenseclaw.jsonVerified end to end
Not tested live: defenseclaw-kiro, in-sandbox login
TermMeaning
Managed tierThe hooks are in a root-owned policy file that user and project settings cannot switch off.
User tierThe agent can edit its hook file, or code it or a repository adds runs beside the hooks. DefenseClaw reports a sandbox whose hooks go silent while the harness is active.
VerifiedThe image passed a hook check against a mock model, and the harness ran end to end in a live sandbox with the daemon: a blocked command stayed blocked and the egress block list held.
UnverifiedThe harness needs a vendor account before its first turn, so no mock model can drive it. sandbox run builds the image, then refuses it.
Not tested liveA sign-in path no live run exercised. The connector's page says which sign-ins sandbox run shares.

openshell.harnesses (Claude Code and Codex when empty) decides which images defenseclaw sandbox setup and a bare defenseclaw sandbox image build build, and which defenseclaw sandbox doctor checks. sandbox setup --harness (repeatable) replaces that list, so name every harness you want. sandbox image build <harness> builds what you name without changing the list, and sandbox run builds a missing image on first use. Every sandbox hook fails closed.

"I want fail-closed on transport failures"

Every listed connector can block a gateway outage: on a per-user install in the strict policy, on a managed host by default. Copilot CLI, Antigravity and Hermes hooks block delivery failures when the effective mode is closed. Hermes 0.21+ blocks a pre_tool_call on hook spawn error or timeout when Setup has written fail_closed: true; it can still run after a killed hook exits non-2 with empty stdout. Earlier Hermes contracts also fail open on timeout. Other agents may still allow a call when the hook process times out or cannot start. The table marks gateway delivery behavior separately from agent-side process failure.

  • Fail-closed in this matrix means the running hook blocks when it cannot reach the gateway. A killed or timed-out pre-tool hook is non-blocking in Claude Code and Codex; a killed pre-tool hook can be non-blocking in Devin. Kiro CLI 2.x can run after hook timeout, and Antigravity has no documented timeout-deny setting. See the enterprise threat model.
  • On a managed standalone host, hooks use the administrator's resolved fail mode, also in observe mode. The default is closed. If the administrator explicitly selects open, a gateway outage can allow the call. Oversized input, a gateway 5xx, an unusable answer and other transport errors still block. See Fail modes.
  • On a per-user install, hook connectors record calls that ran unchecked in hook-failures.jsonl, and the gateway turns those records into a guardrail-degraded alert when it starts again (Reporting). The OpenClaw plugin places its tool-inspection outage warning in the tool result.
  • Cursor writes failClosed: true outside Secure Client in action and observe mode. Observe still allows ordinary calls and records would_block findings; the host flag covers a hook process that cannot answer.

Gateway outage

What each connector's hooks do with a call they can't check because the gateway is not running, is still restarting or refuses the connection before accepting the call. This table is the only place the docs give it per connector, and it is checked against the hook code on every change.

ConnectorPer-user, strict policyPer-user, default or permissive policyManaged standalone, open, default foreign-hook guardManaged standalone, open, foreign_hooks report or allowSecure Client, fail mode closedDEFENSECLAW_STRICT_AVAILABILITY
Claude CodeBlocksRunsRuns; blocks with managed_hooks_only: preserve, which turns the foreign-hook guard onRunsBlocksRead
CodexBlocksRunsRuns; blocks with managed_hooks_only: preserve, which turns the foreign-hook guard onRunsBlocksRead
OpenClawBlocksBlocksBlocksBlocksBlocksNo hooks
CursorBlocksRunsBlocks a new agent process, a session start and every call on Windows; on Linux and macOS, while the gateway service is stopped, a later call of a process the gateway allowed in the last 12 hours does what the next column saysRunsBlocksRead
HermesBlocksRunsBlocks a new agent process or a session start on Linux and macOS; while the gateway service is stopped, a later call of a process the gateway allowed in the last 12 hours, and every call on Windows, does what the next column saysRunsRunsRead
OpenCodeBlocksRunsBlocks a new agent process, a session start and every call on Windows; on Linux and macOS, while the gateway service is stopped, a later call of a process the gateway allowed in the last 12 hours does what the next column saysRuns through machine policy (the default); blocks on the per-user routeBlocksRead
AmpBlocksRunsBlocksBlocksBlocksNot read
GitHub Copilot CLIBlocksRunsBlocks a new agent process, a session start and every call on Windows; on Linux and macOS, while the gateway service is stopped, a later call of a process the gateway allowed in the last 12 hours does what the next column saysRuns through machine policy (the default); blocks on Linux and macOS on the per-user routeRunsRead
AntigravityBlocksRunsRunsRunsRunsRead
DevinBlocksRunsBlocks a new agent process, a session start and every call on Windows; on Linux and macOS, while the gateway service is stopped, a later call of a process the gateway allowed in the last 12 hours does what the next column saysRunsBlocksRead
KiroBlocksRunsRunsRunsBlocksRead
ZeptoClawBlocksBlocksBlocksBlocksBlocksNo hooks
  • Per-user is an install without a managed deployment. An explicitly saved guardrail.hook_fail_mode chooses its outage mode. When that setting is absent, the connector's policy tier (rule pack) decides: the default (balanced) and permissive policies let the call run with one warning and a hook-failures.jsonl record, and the strict policy blocks it. Input over the 1 MiB hook cap is not an outage: in the default and permissive policies the hook sends the gateway a bounded review of it and follows the verdict, and the strict policy blocks it. Every other failure blocks in every tier (Fail modes).
  • Managed standalone is an enterprise host outside Secure Client. Its fail mode is the administrator's hook_fail_mode: closed, the default, blocks the outage on every connector, and the two managed columns give open. A managed connector runs either through vendor machine policy (the administrator-owned defenseclaw-hook --enterprise-managed) or through a hook or plugin the guardian writes into the user's home (the per-user route, for the connectors without machine policy and for one whose machine policy is off).
  • Default foreign-hook guard: with foreign_hooks: remove, the machine policy default, the foreign-hook guard checks each call of a connector it covers against the agent session's hook record, which only the gateway holds. While the gateway cannot answer, the guard blocks that call in every fail mode, before the hook's own fail mode applies. One exception, on Linux and macOS only: when the gateway service is stopped and the gateway had allowed a call of the same agent process and session before it stopped, the guard lets the hook's own fail mode decide, so the call does what the next column says. The column gives both cases. With foreign_hooks: report or allow the guard does not ask the gateway, and the next column applies.
  • Secure Client hooks let an outage run in open, which is also their default when no fail mode is set.
  • Claude Cowork has no row. It is a surface of Claude Code and follows the Claude Code row. The one-time outage warning is never shown inside Cowork; the outage shows in defenseclaw status (Claude Cowork).
  • DEFENSECLAW_STRICT_AVAILABILITY=1 makes an outage block on a hook that reads it, whatever the fail mode.
  • DEFENSECLAW_FAIL_MODE is not read by any 1.0.4 hook: the installed mode is the fail mode (Fail modes).
  • OpenClaw model proxy requests follow the saved per-user outage mode on a classified proxy outage: default and permissive continue to the configured provider with an agent warning, while strict and managed block. ZeptoClaw model proxy requests remain blocked. OpenClaw plugin tool inspection follows the saved mode: default and permissive allow an ordinary call with a warning, while strict and managed block it. Direct protected targets stay blocked.
  • status --json and verify --json report as effective_hook_fail_mode the mode a managed connector's hook works in. That is the administrator's resolved value, except on the per-user route of a standalone install, where the Amp plugin, OpenCode and Copilot CLI (on Windows the Amp and OpenCode plugins) read closed, as their row in the gateway outage table shows. A connector on vendor machine policy reads the administrator's value, and the guard column still applies to it.

Refusals that are not an outage block in every fail mode: an unapproved hook the foreign-hook guard finds, an invalid managed runtime, a refused credential (outside Secure Client, where an HTTP 401 or 403 follows the effective fail mode), and, on a managed standalone host, input over the size limit and an answer DefenseClaw cannot use. See Fail modes.