The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references official pre-built binaries and Docker images hosted on Grafana's public GitHub repositories (github.com/grafana/assistant-cli).
[COMMAND_EXECUTION]: Includes commands to manage system services, such as grafana-assistant tunnel daemon install, which configures the tool to run as a persistent background process.
[COMMAND_EXECUTION]: The assistant tunnel feature allows a remote Grafana Assistant service to execute terminal commands on the local machine when the terminal tool is enabled. The tool documentation notes the existence of deny-lists for dangerous commands.
[DATA_EXFILTRATION]: The assistant tunnel feature provides a mechanism for a remote service to read files from the local filesystem within specified project paths. It includes built-in blocks for sensitive directories like .ssh, .aws, and .env.
[CREDENTIALS_UNSAFE]: The documentation describes standard practices for managing service account tokens using environment variables (GRAFANA_SA_TOKEN) or configuration files. No hardcoded credentials or secrets were found in the skill text.
[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection as it processes untrusted data from external sources (logs, metrics, traces).
Ingestion points: Ingests data from Grafana datasources and local project files via the tunnel.
Boundary markers: Employs configurable allow/deny lists for both filesystem paths and terminal commands to limit the scope of remote actions.
Capability inventory: The tool can execute subprocesses via the terminal tool and read local files via the filesystem tool.
Sanitization: Uses internal block-lists for sensitive files and dangerous shell commands.

grafana-assistant-cli