What TokenEconomizer is
TokenEconomizer is a Python plugin and a paper from Microsoft Research, Microsoft Security and UC Berkeley, published under the Apache 2.0 licence. Its idea is that a coding agent wastes tokens in the same ways every time it works in the same repository, and that the sessions it already finished show where. It names three mechanisms: micro-skills, tool pruning and a failure-loop guard. The Economizer in VibeiDE follows that design for your Claude Code and Codex tasks.
Repository facts: stop failing at the same thing
Agents often rediscover the same fact about a repository by failing at it, task after task: pytest is not on the path but python -m pytest works, a script has another name, a variable must be set first. When three tasks show the same failure and the same way around it, the Economizer keeps it as one short line, and gives it to later agents once a fourth task confirms it is still needed. Agents get at most four facts, about 150 tokens, at the start of a conversation. A fact that stops being true retires by itself, and you can forget one yourself.
Loop guard: one nudge for a stuck turn
A turn that keeps editing and keeps getting the same failure spends tokens without moving. The Economizer first learns how long a turn that ends well runs in your project. A Claude Code turn that goes far past that, still editing and still getting the same failure, gets one note asking for a different approach. If it stays stuck, a line in the task tells you. It never stops a turn. A stuck Codex turn gets no note, only the line for you.
Unused tools: smaller requests
The description of every tool an agent has is sent again with every request. VibeiDE attaches two tool servers of its own, browser tools and task tools, about 3,400 tokens together. Where no task in a project has used one of them in ten prompts over two tasks, new conversations start without it. A prompt that may need it gets it back, and so does Bring back in the Economizer tab. Tools you installed yourself and an agent’s built-in tools are never removed.
One switch, nothing to configure
The Economizer is part of the Brain. Turn the Brain on for a project and it starts reading; turn it off and nothing is read or given. The Economizer tab in the Brain panel shows each fact, what the loop guard has learned and which tools are left out. Its only buttons are Forget and Restore on a fact and Bring back on tools. It reads the agents’ own tool calls, makes no model calls and keeps short facts and counts on your computer.
What it does not promise
Microsoft reports token savings for its plugin from its own benchmark. Those numbers belong to the full plugin, which also uses a model to propose skills and to judge stuck turns. VibeiDE leaves the model-backed parts out, so we state no size of saving. What you can check is in the tab: the facts agents were given, the turns that were redirected and the tools that stayed out of a request.
Sources: TokenEconomizer on GitHub, TokenEconomizer paper.
Related guides
All articles- Claude Code alternatives: 7 tools compared (November 2026)
- OpenCode alternatives: 8 coding agents compared (2026)
- Claude Code review: every way to review code in 2026
- Codex plugin for Claude Code: setup, commands, review loop
- TokenEconomizer: how it cuts coding agent tokens
- Best tools to run multiple AI coding agents in parallel (2026)