Microsoft's TokenEconomizer Plugin Cuts Coding Agent Costs by 18.4–30.9%

Microsoft has published TokenEconomizer on GitHub, a plugin that reduces token consumption in coding agents. It works with four tools: Claude Code, Codex, Copilot CLI and Pi. According to the published benchmarks, the average cost of running agents falls by 18.4–30.9%, and output quality does not drop. The plugin does not change the model itself. It works around the agent and strips out what you would otherwise pay for in vain.

What TokenEconomizer remembers and disables

The plugin learns from the history of previous sessions in a specific repository. The project description names three mechanisms.

First, memory of successful commands. If the agent has already found a working way to build the code, run the tests or check the linter in a project, it does not need to search for that path again on the next task. Every such attempt costs tokens: the agent reads the output, reasons about it and tries another option.

Second, disabling unused tools. Descriptions of connected tools usually go into the context with every request. The more tools are connected, the more tokens go to descriptions, even if the agent has never called most of them in this project. If the session history shows that a tool is not needed, the plugin removes it from the context.

Third, breaking out of loops of repeated errors. It is a familiar situation: the agent runs the same failing command, tweaks the arguments slightly and hits the same error again. Every round is billed. The plugin helps the agent recognize such a loop and change its approach.

Note that all three mechanisms rely on data from a specific repository. The more sessions have accumulated, the more material there is to learn from. The plugin does not impose general "one-size-fits-all" rules for every project.

Where the numbers need checking

The 18.4–30.9% range refers to the average cost of running agents. A spread of almost 12 percentage points shows that the result depends on conditions. The original announcement has no data on which tasks and repositories were measured, which models were used, or how quality was assessed. The phrase "without loss of quality" also needs clarification: which metric, on which set of tasks, and with what tolerance. The answers are worth looking for in the project repository.

There is also an obvious caveat: the benchmarks were published by the vendor itself, and there are no independent measurements yet. A plugin that learns from session history is unlikely to show the same numbers on a new repository with no history. The benefit should grow as sessions accumulate. That conclusion is logical, but the source offers no direct confirmation.

Memory security is a separate question. If the plugin remembers "successful" commands, a team needs to understand where and in what form this is stored, and whether that memory can be viewed and cleaned. Without that, a single bad precedent can stick and keep repeating.

Who should try it

The biggest winners will be teams that run agents on the same large repository every day. That is where repeated commands and surplus tools in the context are most common, and token bills are noticeable. For one-off experiments in a new project, the effect will be smaller.

It is also telling that Microsoft built the plugin for several competing agents at once, including Claude Code and Codex, and not only for its own Copilot CLI. Token savings stop being a single vendor's problem and become a separate layer on top of agents: spending depends not only on the model, but also on how the agent handles context and tools.

The practical advice is simple. Before rolling the plugin out to a team, run it on your own repository and compare two metrics: session cost and the share of tasks solved on the first attempt. If the first drops noticeably and the second holds steady, Microsoft's numbers have held up for your case. The source code and documentation are in the microsoft/TokenEconomizer repository on GitHub.