Agent Firewall

Agent Firewall is a process-level firewall that restricts and audits what autonomous programs, such as AI agents, can access and use.

Example of Agent Firewall blocking a process.

Supported Agents

Agent Firewall supports the securing of any terminal-based agent, including your own custom agents.

Features

Agent Firewall offers network policy enforcement, which blocks domains and HTTP verbs to prevent exfiltration, and writes logs to the workspace.

Agent Firewall also streams audit logs to Coder's control plane for centralized monitoring of HTTP requests.

Getting Started with Agent Firewall

The easiest way to use Agent Firewall is through the agent-firewall module. It can also be ran directly in the terminal by installing the CLI.

Configuration

Agent Firewall is configured using a config.yaml file. This allows you to maintain allow lists and share detailed policies with teammates.

In your Terraform module, install Agent Firewall with minimal configuration:

To use a custom policy, pass it inline via agent_firewall_config, below is an example of minimal configuration for Claude Code module:

For examples of wrapping an agent or process such as Claude Code with Agent Firewall, see the agent-firewall module README.

For a basic recommendation of what to allow for agents, see the Anthropic documentation on default allowed domains. For a comprehensive example of a production Agent Firewall configuration, see the Coder dogfood policy example.

To load the policy from a config.yaml file in your template directory instead, pass it via agent_firewall_config. The module writes the config to the workspace and exposes the resolved path via agent_firewall_config_path, so everyone who launches Agent Firewall manually inside the workspace picks up the same configuration without extra flags. This is especially convenient for managing extensive allow lists in version control.

Configuration Parameters

• allowlist defines the URLs that the agent can access, in addition to the default URLs required for the agent to work. Rules use the format "key=value [key=value ...]":
  • domain=github.com - allows the domain and all its subdomains
  • domain=*.github.com - allows only subdomains (the specific domain is excluded)
  • method=GET,HEAD domain=api.github.com - allows specific HTTP methods for a domain
  • method=POST domain=api.example.com path=/users,/posts - allows specific methods, domain, and paths
  • path=/api/v1/*,/api/v2/* - allows specific URL paths
• jail_type selects the isolation backend. Valid values: nsjail (default), landjail. See Jail Types for a detailed comparison.
• log_dir defines where boundary writes log files.
• log_level defines the verbosity at which requests are logged. Agent Firewall uses the following verbosity levels:
  • WARN: logs only requests that have been blocked by Agent Firewall
  • INFO: logs all requests at a high level
  • DEBUG: logs all requests in detail
• no_user_namespace disables creation of a user namespace inside the jail. Enable this in restricted environments that disallow user namespaces, such as Bottlerocket nodes in EKS auto-mode. Only applies to the nsjail jail type.
• proxy_port defines the port used by the HTTP proxy. Default: 8080.
• use_real_dns uses the host's real DNS resolver inside the jail instead of the built-in dummy DNS server. This allows DNS resolution for non-proxied traffic but permits DNS-based data exfiltration. Default: false.

For detailed information about the rules engine and how to construct allowlist rules, see the rules engine documentation.

You can also run Agent Firewall directly in your workspace and configure it per template. You can do so by installing the binary into the workspace image or at start-up. You can do so with the following command:

When running the binary directly, Agent Firewall reads config.yaml from ~/.config/coder_boundary/ automatically.

Jail Types

Agent Firewall supports two different jail types for process isolation, each with different characteristics and requirements:

1. nsjail - Uses Linux namespaces for isolation. This is the default jail type and provides network namespace isolation. See nsjail documentation for detailed information about runtime requirements and Docker configuration.

2. landjail - Uses Landlock V4 for network isolation. This provides network isolation through the Landlock Linux Security Module (LSM) without requiring network namespace capabilities. See landjail documentation for implementation details.

The choice of jail type depends on your security requirements, available Linux capabilities, and runtime environment. Both nsjail and landjail provide network isolation, but they use different underlying mechanisms. nsjail uses Linux namespaces, while landjail uses Landlock V4. Landjail may be preferred in environments where namespace capabilities are limited or unavailable.

Implementation Comparison: Namespaces+iptables vs Landlock V4

Audit Logs

Agent Firewall streams audit logs to the Coder control plane, providing centralized visibility into HTTP requests made within workspaces—whether from AI agents or ad-hoc commands run with boundary.

Audit logs are independent of application logs:

• Audit logs record Agent Firewall's policy decisions: whether each HTTP request was allowed or denied based on the allowlist rules. These are always sent to the control plane regardless of Agent Firewall's configured log level.
• Application logs are Agent Firewall's operational logs written locally to the workspace. These include startup messages, internal errors, and debugging information controlled by the log_level setting.

For example, if a request to api.example.com is allowed by Agent Firewall but the remote server returns a 500 error, the audit log records decision=allow because Agent Firewall permitted the request. The HTTP response status is not tracked in audit logs.

Audit Log Contents

Each Agent Firewall audit log entry includes:

Viewing Audit Logs

Agent Firewall audit logs are emitted as structured log entries from the Coder server. You can collect and analyze these logs using any log aggregation system such as Grafana Loki.

Example of an allowed request (assuming stderr):