Guardrails for every AI coding agent.
Open source from Cisco AI Defense. Scan skills and MCP servers, block risky tool calls and keep an audit trail for Claude Code, Codex, Copilot, Cursor and 7 more, on macOS, Linux and Windows.
Three commands to your first block
Install. For your user on macOS (Apple silicon (arm64)). Every download is checked against the release checksums.
curl -LsSf https://github.com/cisco-ai-defense/ defenseclaw/ releases/ latest/ download/ install.sh | bash Set up. Finds the agents you have and starts them in observe mode, so nothing is blocked yet.
defenseclaw initTurn on blocking. Switches Claude Code to action mode. The 5-minute quickstart adds a demo rule so you can watch a block.
defenseclaw guardrail mode action --connector claudecode
Release archives, the macOS menu-bar app and checksum steps are on the Download page.
Guided example · Synthetic runtime event
Trace a runtime verdict from event to action
Follow normalization, a bundled rule match, suppressions, severity, and the default pack mapping.
# ~/.defenseclaw/config.yaml (excerpt)guardrail: mode: action # observe mode never blocks judge: enabled: false # the optional LLM judge is off by default # No block_at is set, so the default pack decides: # CRITICAL blocks, HIGH and MEDIUM alert, LOW allows.# ~/.defenseclaw/config.yaml (excerpt)guardrail: mode: action # observe mode never blocks judge: enabled: false # the optional LLM judge is off by default # No block_at is set, so the default pack decides: # CRITICAL blocks, HIGH and MEDIUM alert, LOW allows.
HIGH maps to alert under the default pack
Emit decision record
What DefenseClaw did — and did not do
What it did
- Show the ordered stages that assemble a runtime verdict
- Distinguish the optional judge from deterministic rules
What it did not do
- Evaluate Rego or scanner output in the browser
- Reuse runtime actions as skill or MCP admission actions
11 agents, one policy
Each connector uses the strongest control its agent exposes: native hooks, a policy plugin, or a bridge.
- Claude Code
- Linux: supported
- macOS: supported
- Windows: supported
- Codex
- Linux: supported
- macOS: supported
- Windows: supported
- GitHub Copilot CLI
- Linux: supported
- macOS: supported
- Windows: supported
- Cursor
- Linux: supported
- macOS: supported
- Windows: supported
- Devin
- Linux: supported
- macOS: supported
- Windows: supported
- Antigravity
- Linux: supported
- macOS: supported
- Windows: supported
- Kiro
- Linux: supported
- macOS: supported
- Windows: supported
- OpenCode
- Linux: supported
- macOS: supported
- Windows: supported
- Amp
- Linux: supported
- macOS: supported
- Windows: supported
- Hermes
- Linux: supported
- macOS: supported
- Windows: supported
- Claude Cowork
- Linux: preview
- macOS: supported
- Windows: supported
Supported Preview Not supported
See the support matrixBefore, during and after every tool call
Admission checks, runtime decisions and the evidence trail share one policy and one audit log.
- GuardrailsStart in observe, switch to action, and block risky tool calls before they run.
- Skill and MCP scanningCheck skills and MCP servers before an agent can load them.
- AI and runtime discoveryFind the agents, local models, MCP servers and skills on a machine, and see which ones ran.
- Human approvalPause a high-risk action for a person to approve, where the agent supports it.
- ObservabilitySend every decision over OTLP to Grafana or Galileo, or to a webhook.
- Policy creatorBuild a policy section by section in the browser and copy the YAML into your config.
- Terminal UIAlerts, scans and setup in one terminal app on every OS. A macOS menu-bar app is in preview.
- OpenShell sandboxesPreviewRun a coding agent in an NVIDIA OpenShell sandbox on Linux or an Apple-silicon Mac, with every tool call judged.
Protection your users can’t turn off.
- Admin-owned
- Services, policy and program files belong to the administrator. Users and their agents can’t turn protection off.
- Machine policy
- Hooks installed at machine level for Claude Code, Codex, Copilot, Cursor, and OpenCode with the managed plugin.
- Your MDM
- Recipes for Intune, Jamf, Iru (formerly Kandji), Workspace ONE, ConfigMgr and Linux configuration management.
See the decision in context
Walkthroughs that connect configuration to the interception point, the outcome and the audit evidence.
- Stop Claude Code from running a destructive commandProve the action never reaches the disk.
- Catch a prompt injection on CodexCombine deterministic rules with an optional judge.
- Approve risky tool calls before they firePause HIGH findings and return the operator’s decision.
- Pin local observability in under 60 secondsFollow one event through metrics, logs, traces, and audit.
Put a guardrail around your first agent in five minutes.
No LLM key is required for deterministic runtime rules or static scanner checks.