Token Economizer
A context optimization tool released by Microsoft that analyzes repositories and conversation history to automatically identify and remove tool definitions that are never used, reducing the fixed token overhead of coding agents.
Timeline (last 90 days)
All repository information learned by Token Economizer is stored in a local folder and is not uploaded to the cloud, and the project itself is open source.
Failure-loop intervention uses a separate "judge model" to determine whether the agent is stuck in a repetitive loop, and the judge model only gives hints rather than forcibly terminating the task.
Token Economizer's benefits depend heavily on repeated sessions on the same repository: the longer it is used and the more it learns, the more it saves; when frequently switching between unrelated projects, the effect is noticeably reduced.
Each agent learns independently, and what is learned on Claude Code is not automatically transferred to Codex.
The benefits of micro skills and tool pruning are not linear: savings in the first session are close to zero or even slightly negative, and they usually only become stable after several rounds, or even a dozen or more rounds, of similar tasks.
Microsoft has open-sourced a free plugin called Token Economizer, used to reduce the token cost of AI coding agents.
In joint testing by Microsoft and UC Berkeley, Token Economizer reduced the average cost of AI coding agents by 18.4% to about 31%.
The test covered 6 repositories, 1777 SWE-bench tasks, and 5 agent-and-model combinations, with a per-run cost of about 200 USD.
Among the 5 agent-and-model combinations, after enabling Token Economizer, the task success rate rose rather than fell in 4 of them and declined slightly in 1.
In all 30 "repository × agent-model" combinations, enabling the plugin was cheaper than not enabling it.
8 more timeline events
All Facts (18)
All repository information learned by Token Economizer is stored in a local folder and is not uploaded to the cloud, and the project itself is open source.
50%UnverifiedThe micro skills feature is disabled by default and must be enabled manually.
50%UnverifiedFailure-loop intervention uses a separate "judge model" to determine whether the agent is stuck in a repetitive loop, and the judge model only gives hints rather than forcibly terminating the task.
50%UnverifiedToken Economizer's benefits depend heavily on repeated sessions on the same repository: the longer it is used and the more it learns, the more it saves; when frequently switching between unrelated projects, the effect is noticeably reduced.
50%UnverifiedEach agent learns independently, and what is learned on Claude Code is not automatically transferred to Codex.
50%UnverifiedThe benefits of micro skills and tool pruning are not linear: savings in the first session are close to zero or even slightly negative, and they usually only become stable after several rounds, or even a dozen or more rounds, of similar tasks.
50%UnverifiedIn joint testing by Microsoft and UC Berkeley, Token Economizer reduced the average cost of AI coding agents by 18.4% to about 31%.
50%UnverifiedThe test covered 6 repositories, 1777 SWE-bench tasks, and 5 agent-and-model combinations, with a per-run cost of about 200 USD.
50%UnverifiedAmong the 5 agent-and-model combinations, after enabling Token Economizer, the task success rate rose rather than fell in 4 of them and declined slightly in 1.
50%UnverifiedIn all 30 "repository × agent-model" combinations, enabling the plugin was cheaper than not enabling it.
50%UnverifiedToken Economizer's savings figures are net gains, having already deducted the extra model-call overhead required for the plugin itself to run.
50%UnverifiedToken Economizer acts directly on mainstream AI coding agent tools such as Claude Code and Codex, without switching agents or using cheaper models.
50%UnverifiedToken Economizer contains three core mechanisms: micro skills, tool pruning, and failure-loop intervention.
50%UnverifiedIn Microsoft's own tests, Token Economizer delivered a cost reduction of 18% to 31%.
50%UnverifiedRTK handles noise on the input side, Caveman handles redundancy on the output side, and Token Economizer handles historical baggage in the context; the three do not replace one another, and only when used together do they cover the complete chain of waste.
50%UnverifiedToken Economizer is a tool developed by Microsoft that learns from the user's repository and past sessions, identifies which tool definitions are almost never called, and trims them from the context.
50%UnverifiedMicrosoft has open-sourced a free plugin called Token Economizer, used to reduce the token cost of AI coding agents.
50%UnverifiedMicrosoft quietly open-sourced the Cloud Code efficiency plugin Token Economizer, specifically to help Agents reduce Token consumption, supporting Cloud Code, Codex, GitHub Copilot CLI and others.
50%