wtf is token-saviour?

token-saviour is an agent skill that helps AI coding assistants like Claude Code and Codex spend fewer tokens by picking the right specialized tool for each kind of task, instead of always reading whole files or dumping raw command output into the conversation. The idea is that token cost breaks down into four separate layers: reading code, reading command output like test results or git logs, writing prose replies, and writing generated code, and a different existing tool handles each layer best. Based on a benchmark the author ran across nine different tools, this skill routes code reading tasks to a tool called serena, command output tasks to a tool called rtk, prose writing to a tool called caveman, and code generation to a tool called Ponytail. According to the measurements shared in the README, combining serena, rtk, and caveman together cut token usage by about seventy percent on a comprehension focused task, with the code reading tool serena responsible for most of that saving on its own. Rather than requiring all four tools installed, the skill classifies each task by which layer it stresses, then routes to the matching tool with specific commands, falling back to plain reading and searching tools if the specialized tool isn't installed. It also refuses certain combinations the author found counterproductive, like using two code reading tools at once, or using the command output tool for a comprehension task, and it announces which tools it used for a given task. You can install token-saviour as a Claude Code plugin, as a personal skill copied into a config folder, or as a Codex plugin, and it also ships an always-on instructions file that works in any agent that reads a plain instructions file with no extra setup. The project is released under the MIT license. This suits developers who use AI coding agents heavily and want to reduce token spend and cost without giving up capability, particularly for large codebases where reading entire files repeatedly gets expensive.