The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The td_execute_python tool allows the agent to execute arbitrary Python code within the TouchDesigner process. This environment has unrestricted access to the local filesystem and network under the user's permissions.
[DATA_EXFILTRATION]: The td_get_screen_screenshot and td_get_screenshots tools provide the capability to capture the actual desktop screen and specific software views. This presents a high risk for the exposure of sensitive information visible on the user's monitor.
[COMMAND_EXECUTION]: The scripts/setup.sh script executes several system commands to verify the environment, download external components, and modify the agent's internal configuration.
[EXTERNAL_DOWNLOADS]: The setup script downloads a binary plugin (twozero.tox) from an external domain (404zero.com). This domain is not listed as a trusted vendor, introducing supply chain risk.
[DYNAMIC_EXECUTION]: The skill uses dynamic execution patterns, particularly through td_execute_python, to perform multi-step logic, wiring, and operator creation based on agent-generated scripts.
[DYNAMIC_CONTEXT_INJECTION]: SKILL.md contains a shell command executed via the ! pattern (!nc -z 127.0.0.1 40404...) that runs at skill load time to perform a network health check.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process data from various untrusted external sources, including WebSockets, MQTT brokers, and HTTP servers (references/external-data.md). Malicious instructions embedded in these data streams could be parsed by the agent and executed via its powerful Python and system tools.
[PRIVILEGE_ESCALATION]: The automated setup process modifies the agent's core configuration file (config.yaml) to register new servers and change timeout settings, which is a significant modification of the agent's execution scope.

touchdesigner-mcp