One headline turned a paper into "Claude's usage limits are finally enough": Microsoft's token-saving plugin gained 4,000 stars in a week

microsoft/TokenEconomizer is the companion plugin to a paper from Microsoft Research and UC Berkeley. It plugs into Claude Code, Codex, Copilot CLI, and Pi, remembers the pitfalls you hit in the same repo, and spends fewer tokens next time. The paper came out in October and gathered only about two hundred stars in two weeks. On Oct 28 a YouTube channel framed it as "fixing Claude's usage limits", and AI media and big X accounts kept reposting it, adding 4,278 stars in a week, reaching 4,834 stars by 11-06. Below: what it is, where the stars came from, how crowded this niche is, and what paths are open if you want a similar project to gain stars.

One video, one rewritten headline

On 2026 Oct 28 at 3 UTC in the afternoon, the YouTube channel AI LABS, with 179k subscribers, posted a video titled "Microsoft's New Tool Finally Fixed Claude's Usage Limits". The first line of the description: "Claude burns through its usage quickly because Claude is wasting tokens." In the video, the plugin is installed on a demo project and turns off 276 tools that were never used; each message drops from about 34,300 tokens to 24,400 (as relayed by commenters). By 11-06 the video had 56,766 views and 910 likes (YouTube, 11-06).

The paper's own wording is "cache-aware cost reduced by 18.4–30.9%", and Aaron J. Li wrote the same thing in his X post on Oct 26, which gained only 5 stars the next day. Recast as "Claude's usage limits are finally enough", the same claim gained 181 stars the following day. The comments were also telling: some said "I trust Microsoft to fix things", some asked "aren't /doctor and the new nested projects already solving this?", and some asked whether it works with the Claude desktop app.

What it is

A plugin that runs outside the coding agent, doesn't swap the model or change the agent, and works through the agent's own hooks and settings. The core idea: you use the same agent in the same repo every day, and a lot of the money is spent repeatedly. It handles three things:

• Remember how the repo works: e.g. it needs python -m pytest not pytest, or a build has to run before some command. A lesson only gets told to the agent at session start once it's confirmed across three sessions in a row; if it's later overturned, it's deleted.
• Cut tools nobody uses: every loaded tool definition is resent with every request. Tools unused across several sessions in a row stop being loaded; one command brings them back.
• Stop dead loops: when the same error keeps recurring, a judge model checks whether the agent is stuck; if so, it adds a line telling it to try another approach. If it's still stuck, it tells you, but never stops things on its own.

The math uses real billing: the cached portion is priced at cache rates, and the plugin's own model calls are counted too. The paper tests it on 277 SWE-Bench tasks across 5 agent and model combinations: cost drops 18.4–30.9%, 30 of the cells are cheaper across the board, and pass rate rises slightly for the 4/5 combination (README).

How the stars came in, day by day

(Daily GitHub snapshot, Beijing time; 292 forks.)

The starting point was neither the paper crowd nor HN (11-04 submission, 2 points), but YouTube AI tutorial channels and AI news accounts on X, amplified by MarkTechPost; Reddit came later. Daily star growth has already dropped from 1,219 to 66, so the first wave is over.

Why this one

The token-saving niche was already crowded, and others posted bigger numbers:

Among a pile of personal projects each claiming 90% savings, what it has is something different: credibility. An official Microsoft org, with Berkeley's Ion Stoica among the authors; there's a paper and reproducible experiments, the numbers are conservative, and it's measured against real bills with its own overhead included. Nearly every viral headline is "Microsoft + save 30% + fix Claude's limits"; nobody is talking about the cross-session learning mechanism. The mechanism gives evaluation videos and comment sections something to discuss, but the traffic comes from the headline.

It also covers Claude Code, Codex, and Copilot in one plugin, so users of all three will share it.

What it doesn't do well

v0.1.0, with plenty of problems; dozens of issues were opened in a single week (GitHub, 11-06):

• Most problems are on Windows: command lines too long and erroring, temp directories that can't be deleted, hooks launching Python twice per start, and within 8 hours a system service's memory ballooned to 23.6 GB (#144).
• Dead-loop detection false positives: Playwright tests passed, the agent read source code, and edited text containing "assert" or "failed" was all treated as being stuck (#128, #151).
• It wastes money itself: under Claude Code, one Haiku call from the plugin pulls in every MCP server, roughly 570k effective tokens (#87); someone also said "I've used it a lot and can't see what the plugin is doing" (#113).
• Sessions from before installation aren't learned from: comments and issues ask for "read the history in ~/.claude first when installing", and an external PR already exists for it (#130).

But external developers are filling these gaps in the main repo: 11-03 to 11-06, external PRs arrive every day, with people adapting it to Factory Droid, Grok Build, Antigravity CLI, and Hermes. On GitHub I found no similarly named or described repo built from scratch after 10-25. Building another one to fill its gaps would most likely mean doing contract work for it.

If it were you

C: You can learn a playbook from it, but it's not worth building on its own. Its stars mostly come from the Microsoft and Berkeley names, which you don't get; the parts it does poorly are being filled by the community in the main repo; and in the "save tokens" space, four or five projects already sit at a hundred thousand stars.

Who it suits: developers with an agent tool whose effect can really be measured, willing to put effort into a reproducible evaluation; people in a school or research group who were going to publish a paper anyway.

Who it doesn't suit: anyone wanting to build another "save tokens / agent memory" plugin (caveman, claude-mem, and rtk already hold those spots, and TE holds the "reliable" position); anyone wanting to make a Windows or Cursor version of TE (the community is already sending those to the main repo); anyone with no channel into the AI YouTube channels and X news accounts.

How to do it: if you insist on building in this space, the best shot is a position neither it nor its rivals hold: an open-source evaluation that scores every token-saving tool against real billing. Run the same tasks with caveman, rtk, claude-mem, and TE each, compute cost on cached billing, count each tool's own overhead, and publish the results as a table people will share. rtk's "it saves tokens but not money" criticism shows people want exactly this, and evaluation projects naturally get shared by the users of every tool being tested. The other route is to borrow its playbook for a different problem: "paper or reproducible experiment + a plugin installable in one command + a GIF + one conservative number", with a headline that puts the user's actual pain in their own words.

How to get users: reverse TE's sequence. Don't count on HN or the paper crowd at first; send the project and a line like "does X tool actually save money?" straight to AI tutorial channels like AI LABS, AI media like MarkTechPost, and AI news accounts on X, since they're short on topics. Put a results table and a one-line install at the top of the README.

When it's too late: Claude and Codex usage-limit anxiety is still flaring, and interest in token saving rides on that. Once Anthropic or OpenAI build tool trimming and cross-session memory into the product itself, the topic goes cold; this needs to ship by roughly the first half of 2027.

What would mean you've misread it: (1) Claude Code or Codex officially ships automatic tool disabling based on usage or repo-lesson memory, and third-party plugins across the category slow down; (2) your evaluation gets shared by one or two AI accounts but doesn't reach a few hundred stars in a week, meaning people only care about "made by Microsoft" headlines, not who actually saves money; (3) conversely, if TE's daily growth picks up again in mid-to-late November (say, it hits GitHub trending), demand is bigger than estimated here and the category has room for one more entrant.