Tools and parameters
Define which agent may use which tool and under what conditions.
Security for AI agents
Let Claude Code, Codex and other agents work with your repositories, cloud and databases. Every tool call goes through your policy, a human approves sensitive steps, and everything ends up in an audit log that cannot be rewritten unnoticed.
Join the pilotHow it works
A prompt is not a security boundary. AgentWall enforces what an agent can actually do, independently of what a model or untrusted data suggests.
How it works
Every request ends with one of three outcomes:
Allow routine safe actions.
Block forbidden requests.
Approve sensitive steps with a human.
Works with your agents
For Claude Code, Codex and in Docker Sandboxes, AgentWall locks the managed configuration and, together with the sandbox, the only path to tools and to the network goes through your policy. For Gemini CLI, OpenCode, VS Code and Hermes Agent, it sets the configuration and watches for drift. Installation is one command, and you pair the device with a code in the browser. The model and the subscription stay yours, local models included.
What AgentWall protects
Define which agent may use which tool and under what conditions.
A risky action waits with its exact parameters, and approval unlocks only that one action.
AgentWall holds keys and tokens, the agent never sees them. For company MCP servers configured in the control plane, they stay on the device only. Every tool decision and approval is recorded.
In isolated mode, the agent’s outbound connections reach only allowed destinations. The internal network, localhost and cloud metadata stay out of reach, even behind an allowed domain.
Your control plane, run by you or hosted by AgentWall, issues policies as a signed bundle. Devices reject older or forged policies and allow nothing without a valid one.
See which MCP servers your devices use and decide for each one: through AgentWall, allow, or block. AgentWall reports an added server or a changed setting, and a foreign server can be moved to a reversible quarantine.
Verified, not promised
In isolated mode, a confined agent tries to reach the protected system directly with curl, Python and PowerShell, with and without a token. Every attempt stops at the network boundary, also after a restart.
An allowed destination is reachable only through AgentWall. The proxy refuses the protected system and unknown servers and records every decision.
FAQ
AgentWall is a policy, human-approval and audit layer between AI agents and company systems. It decides on every tool call (MCP) and, in isolated mode, every outbound connection an agent makes before it runs: allow it, block it, or send it to a human for approval.
No. AgentWall uses the sandbox you already have (for example the client's own sandbox or Docker Sandboxes) and puts an MCP gateway and egress proxy with your policy on the only way out. You keep your agent and your model.
Claude Code, Codex, Gemini CLI, OpenCode, VS Code (Copilot) and Hermes Agent, local models included. For Claude Code, Codex and in Docker Sandboxes, AgentWall locks the managed configuration and, together with the sandbox, the path to tools and to the network goes only through your policy. For Gemini CLI, OpenCode, VS Code and Hermes Agent, it sets the configuration and watches for drift.
AgentWall does not detect prompt injection in text. It limits what an agent can actually do. Even if a model or poisoned data steers the agent towards a forbidden action, AgentWall blocks it or puts it in front of a human for approval.
A sensitive action waits in a queue with its exact parameters. Approval unlocks exactly that one action with those parameters, nothing else. Decide on the device or in the web console, with notifications via webhook.
Records form a hash chain and the database refuses to modify them. A signed checkpoint, verifiable offline, reveals even tampering by a database administrator.
You can run the control plane that issues signed policies yourself or use the hosted AgentWall control plane. Devices reject older or forged policies.
AgentWall pilot programme
We are looking for companies that already use Claude Code, Codex or other agents with access to repositories, cloud or databases, as well as cybersecurity partners and integrators.
We will send you a confirmation e-mail. No spam, pilot information only. How we process personal data.