Awsmux: Multi-Account AWS CLI Speeds AI Agents Up to 5.4x
Awsmux runs AWS CLI commands across hundreds of accounts in parallel, cutting AI agent costs and latency up to 5.4x with an approval-gated safety layer.
Awsmux is an open-source Go tool that fans out a single AWS CLI command across hundreds of accounts and regions in parallel, using 100 concurrent workers by default, verifying each identity via STS, and merging results into one stream. Mutating commands are gated behind an immutable plan/approve/apply workflow, with a one-time approval token bound to the plan's sha256 hash — so even an AI agent holding admin credentials can't alter or replay an approved action.
The tool ships an MCP server exposing five structured tools (target listing, plan creation, execution, etc.) for clients like Claude Code, Claude Desktop, and Cursor. In a published 150-session, three-arm benchmark using identical Claude Opus 4.8 agents and prompts, the awsmux arm beat a raw-shell aws CLI arm in every cell: 1.3x-2.9x cheaper, 2.3x-5.4x faster, and up to 7.4x fewer output tokens, holding a flat 4 turns regardless of fleet size while the CLI arm grew from matching that at 10 accounts to 10.5 turns at scale.
For engineers managing large AWS fleets, this translates into measurable cost and latency savings for both manual operations and agent-delegated tasks, plus a hard safety boundary around destructive actions. A bundled 100-account LocalStack sandbox lets teams test the full workflow with zero real credentials.
This synthesis was produced from its source by AI; there is no human editor or manual review step. How we work