The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the agent to perform an extensive series of shell operations on the host to audit a Docker image before use. This process includes pulling the image, extracting its root filesystem into a temporary directory, and running complex find and grep commands to identify potential risks like hardcoded credentials, native binaries, or lifecycle hooks. Each stage requires explicit user approval.
[EXTERNAL_DOWNLOADS]: The skill fetches a Docker image from the GitHub Container Registry (ghcr.io), which is a well-known service. Security is maintained by pinning the download to a specific, immutable SHA-256 digest (sha256:751a0d32acd2d6b1da6ad9ac67987fbd1ff36ce26b7160014d8605f18b7907b3), preventing the agent from pulling untrusted or updated versions without further review.
[REMOTE_CODE_EXECUTION]: The skill registers and executes an MCP server via docker run. It implements rigorous security controls, including dropping all Linux capabilities (--cap-drop ALL), disabling container networking (--network none), and preventing privilege escalation (--security-opt no-new-privileges). These measures ensure that the gateway code executes in a highly restricted environment.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from its own MCP tools, which could potentially contain untrusted instructions if the gateway were compromised.
Ingestion points: Data enters the context via outputs from tools such as gateway_health, gateway_readiness, and gateway_chat.
Boundary markers: The skill includes explicit warnings to treat a zero exit code only as transport evidence and to maintain a deterministic, credential-free mode by default.
Capability inventory: The skill's capabilities are limited to the nine specific MCP tools provided by the server, running in a network-isolated container.
Sanitization: The skill relies on the structured nature of the MCP tool outputs and provides instructions to report blockers exactly as returned, preventing the agent from interpreting raw data as instructions.

unified-ai-gateway