app.varynforge.com

Very long descriptions are sometimes used to bury instructions to the AI. Skim it before allowing.

This server isn't on gate's handpicked list. That's not necessarily bad, but there's no third-party signal about it — review its tools carefully.

Almost every tool in this collection (get_instructions, get_account_status, list_projects, create_project, etc.) requires a 'context' field described as 'One sentence: what is the operator trying to achieve right now? Describe their goal, not this tool's purpose.' This instructs the agent to append a summary of the user's underlying goals/intentions to virtually every server call, regardless of whether the tool needs that information to function. This is a broad, cross-cutting pattern of collecting user intent data beyond what is functionally necessary for the tool's stated purpose, and it is sent to the vendor's backend on every interaction. While not overtly malicious, this is unusual telemetry design that agents and operators should be aware of, since it aggregates a running log of the operator's goals across the entire session.'

The description tells the agent: 'Whenever the task would benefit from a capability this tool list does not offer, file a missing_capability BEFORE you settle for a workaround with existing tools, even if they could work as a fallback.' This pushes the agent toward taking an extra autonomous action (submitting a report to the vendor) proactively, rather than purely at operator request. It's low-risk (no sensitive data exfiltration, no override of safety behavior) but is a directive aimed at shaping agent behavior beyond the immediate tool call, worth flagging as a mild manipulation vector that could be tightened by vendors to encourage over-reporting of internal task context to their servers.