The Agent Skills Directory

[DATA_EXFILTRATION]: The skill performs network operations to api.sarvam.ai to process speech and language tasks. These requests involve transmitting user-provided data, such as text or audio recordings, to the vendor's API for processing.
[EXTERNAL_DOWNLOADS]: The installation guide (references/install.md) instructs users and agents to install the sarvam-mcp package using pip or uvx. This package is a dependency provided by the skill's author to enable MCP functionality.
[DATA_EXPOSURE]: The skill and its documentation reference the use of ~/.sarvam/credentials for storing and retrieving API keys. Accessing this specific file is part of the standard authentication flow for the Sarvam platform.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a processing engine for untrusted external data, including audio recordings, text inputs, and localization files. This creates a potential surface for indirect prompt injection if the processed content contains hidden instructions.
Ingestion points: User-provided audio files, text strings for translation, and structured data files (JSON, YAML, etc.) in localize tasks.
Boundary markers: The instructions do not specify explicit delimiters or "ignore instructions" wrappers for the data being sent to the LLM or speech tools.
Capability inventory: The skill has the ability to read local files (audio/text), write output files to the local system (defaulting to ~/Desktop), and make external API calls to api.sarvam.ai.
Sanitization: No specific data sanitization or validation steps for input content are described in the instructions.

sarvam-mcp