Execute computer actions on a browser session. Pass a single action for simple operations (e.g. one click or one screenshot), or pass multiple actions to batch them into a single request for lower latency (e.g. click, type, press_key in one call). Use sleep actions between steps when the page nee...
AI agents invoke computer_action to trigger actions in Kernel MCP Server. What it does depends on the arguments the agent supplies, and its effects often reach beyond the immediate call — builds kicked off, notifications sent, workflows started.
This tool executes arbitrary browser and computer actions (clicks, typing, key presses, clipboard writes, CLI reads) whose effects depend entirely on the arguments provided. While it does not delete data or move money directly, it can trigger any web-based action an agent instructs it to perform, including form submission, navigation, data entry, and command execution.
From the tool's definition Tool description states 'Execute computer actions on a browser session' and lists action types including 'click_mouse', 'type_text', 'press_key', 'scroll', 'drag_mouse', 'write_clipboard', 'read_cli' which are explicit browser automation and system input…
Documented attack patterns abuse exactly the kind of access computer_action gives an agent:
PolicyLayer is an MCP gateway — it sits between your AI agents and Kernel MCP Server, and nothing reaches the server without passing your rules. This is the rule we recommend for computer_action:
{
"version": "1",
"default": "deny",
"tools": {
"computer_action": {
"limits": [
{
"counter": "computer_action_rate",
"window": "minute",
"max": 10,
"scope": "grant"
}
]
}
}
}computer_action stays usable, but rate-capped — a runaway agent can't fire it dozens of times a minute. Everything else on the server is denied unless you say otherwise.
Free to start. No card required.
Execute computer actions on a browser session. Pass a single action for simple operations (e.g. one click or one screenshot), or pass multiple actions to batch them into a single request for lower latency (e.g. click, type, press_key in one call). Use sleep actions between steps when the page needs time to react (e.g. after a click that triggers navigation or animation). IMPORTANT: Always include a screenshot as the last action so you can see the result of your actions. Action types: click_mouse, move_mouse, type_text, press_key, scroll, drag_mouse, set_cursor, sleep, write_clipboard, read_clipboard, screenshot, get_mouse_position. screenshot, get_mouse_position, and read_clipboard return data, so they must be the last action if included. It is categorised as a Execute tool in the Kernel MCP Server MCP Server, which means it can trigger actions or run processes. Use rate limits and argument validation.
Register the Kernel MCP Server MCP server in PolicyLayer and add a rule for computer_action: allow, deny, rate-limit, or require approval. Point your MCP client at the PolicyLayer proxy URL and the rule is enforced on every call, before it reaches Kernel MCP Server. Nothing to install.
computer_action is a Execute tool with high risk. Execute tools should be rate-limited and have argument validation enabled.
Yes. Add a rate_limit block to the computer_action rule in your PolicyLayer policy. For example, setting max: 10 and window: 60 limits the tool to 10 calls per minute. Rate limits are tracked per agent session and reset automatically.
Set action: deny in the PolicyLayer policy for computer_action. The AI agent will receive a policy violation error and cannot call the tool. You can also include a reason field to explain why the tool is blocked.
computer_action is provided by the Kernel MCP Server MCP server (kernel/kernel-mcp-server). PolicyLayer sits as a proxy in front of this server to enforce policies before tool calls reach the server.
Start from Kernel MCP Server, add the rest of your stack, and see everything your agents can call. Then put policy on all of it.
Free to start. No card required.
16 Kernel MCP Server tools catalogued and risk-classified — across an index of 43,000+ MCP servers.