Built for teams running local or airgapped models who need browser automation, desktop control, and code workflows without sending credentials to cloud APIs. Ships 52 capability packs as typed JSON tools that 7B–30B models can call reliably: browser sessions, git operations, vision tasks, and multi-step pipelines all wrapped behind single MCP calls. The pitch is cost and control: a six-step code edit loop runs for $0.07 on gpt-oss-120b versus $0.30+ on Sonnet, and everything stays in your VPC. Includes routing memory, intent decomposition, and a management UI. If you're optimizing for small model success rates or can't let an LLM see your SaaS tokens, this is the stack.
claude mcp add --transport stdio tosin2013-helmdeck uvx helmdeck