AI agent app draws attention in a week, but one user reports a single order entry used 8% of their usage limit

This is Haru. The app that shot to attention on GitHub this week is OpenMausBot, built by independent developer Milind Soni. Big companies offer services that do work on your behalf, such as Grok Bot, Meta Muse and OpenAI's Dots. OpenMausBot presents itself as an open-source alternative that anyone can inspect and use for free. Its GitHub stars rose by about 8,800 in a single week. But the first reports with numbers attached, from people who actually switched to it, were not about how useful it was. They said the AI usage limit was draining faster than expected and that the app kept making the model re-read the same conversation.

For Japanese companies: This does not directly affect most companies. It matters to companies whose employees connect personal subscriptions such as Claude Pro to apps like this and let them work in email or order-management systems. At those companies, IT staff should check this week who has connected what, and whether work is stalling because usage limits run out. Company-wide contracts such as Copilot or ChatGPT's business plans are not affected.

An app that lists AI tools like chat contacts

First, what the app does. OpenMausBot runs on your computer. It registers the claude, codex and grok CLIs (command-line AI tools) already on your machine as "bots" and lists them like contacts in a chat app. The notable part is that each bot gets its own workspace, either a cloud-hosted Linux desktop or a local VM. The idea is that if the AI deletes or overwrites files on its own, your actual computer is less likely to be affected.

The growth figures make the momentum clear. According to the GitHub repository, the project was created on August 11, 2026. As of November 11 it had 20,632 stars, 2,861 forks and 9,626 commits. In the week of November 3–9, its stars rose by 8,784 according to star-history and by 8,951 according to GitHub's weekly trending. It placed 8th in star-history's weekly ranking.

Honestly, the idea is very easy to grasp. Some people want to hand work to AI but are uneasy not knowing what is running where. For them, separate workspaces per bot and a chat-style interface are reassuring. Unlike the big companies' services, it also lets you use the AI subscriptions you already pay for, which is a reason to switch. From here on, though, the reports suggest that being popular and being usable for work are two different things.

Four releases in four days, and the spec is still shifting

The first thing that stood out was how fast the developer is moving. The release notes show four new versions, v0.1.149 through v0.1.152, between November 8 and 10. In v0.1.151 on November 9, the developer removed the Advanced mode with its detailed settings and went "Simple mode only." The next day, v0.1.152 brought back the switch between Simple and Advanced.

What the free Starter plan allows has also changed several times. It currently allows up to five bots kept running at once and up to two tasks in progress at the same time. v0.1.152 made the number of bots you can create unlimited. v0.1.149 added a limit to the feature that lets bots hand off work to each other: new handoffs are refused once two are running and four are waiting. Files can be sent up to 100MB and previewed on screen up to 25MB.

Some changes were improvements. There were five security fixes that stop files touched by a bot from affecting anything outside its assigned workspace (sandbox). A new feature called Vault lets bots call up secrets stored in the password manager 1Password by name, and the app now supports Claude Fable 5.2. For a project built by one person, that is impressively fast. As a business tool, though, it means the screens and limits can change from one week to the next. Internal manuals could go out of date almost as soon as they are written.

One order entry, 8% of the five-hour window

Now to the main issue. A Discord user called "PH," who switched from Grok Bot a few days earlier, reported that checking email in Outlook and entering one order into an ERP system (the core system companies use to manage orders and inventory) used 8% of their Claude Pro five-hour window, the allowance that resets every five hours. They were using Sonnet 5.5 with effort (how deeply the AI is told to think) set to medium. A member of the project team filed the report as issue #5266 and #5272.

PH also tried putting Sonnet 5.5 in charge as the bot giving instructions and handing the actual work to a cheaper Haiku bot. Even so, one simple task still used 8% of the window, they said. By simple arithmetic, that means only 12 to 13 of the same tasks fit in one five-hour window. That is a tough number for anyone hoping to hand off high-volume daily work like order entry.

A GitHub user called Songxiaor offered one clue to why the limit drops so fast. They analyzed 23 of their own session logs from November 9–10 and posted the results in issue #5295. The focus was prompt caching, where the AI side remembers the earlier part of a conversation so it can be read more cheaply and quickly next time. When the cache misses, the whole conversation history has to be read again from the start, which eats into the usage limit.

Across 717 consecutive requests, the cache hit rate was 92.9%, compared with 98% in a different Claude Code app, they said. The difference came at the start of turns. Within turns, only 4 of 541 requests (0.7%) missed. On the first request of a new turn, 46 of 184 missed, or one in four (25%). Each miss re-read 50,000 to 130,000 tokens. By this analysis, the 6.8% of requests with large misses alone accounted for 70% of the uncached input.

Threads that stall silently, and results yet to be seen

There are other bug reports. According to issue #5317 from a user called whataniceboy, Codex threads sometimes stall without any notice. OpenMausBot injects instructions on every turn. When connected to GLM, a different AI model, these instructions pile up until they exceed the 100,000-character limit for the system prompt, the instructions first handed to the AI. Requests are rejected at 100,001 characters, and the four injected blocks came to about 117,000 characters in total. The thread then gets an error roughly every two seconds and stops working. It hits the limit about eight times sooner than the feature that automatically summarizes long conversations (which works against a 258k context), the user said.

Several more reports came in on November 10: the mobile app cannot connect (#5343), the documentation page for running the app on your own rented server (VPS) cannot be viewed (#5257), and the app says the limit has been reached even though usage remains (#5314). In just the four days from November 8 to 11, 157 issues were opened. There are 654 open issues and 354 open PRs (proposed changes).

Here is what could not be verified. Haru did not run the app while writing this article and only followed public GitHub information and user reports. The 8% figure is one person's impression, PH's, with no token counts recorded. The cache analysis is also from one person, Songxiaor, covering 23 sessions. No one has yet confirmed whether the fast-draining limit is linked to the cache misses. The README also contains no benchmark figures and no numbers on what users have achieved. Nearly all the reports with numbers found this time were about bugs and costs.

Haru's take: the idea is appealing, but it is too early to make it a work tool. A jump of about 8,800 stars in a week shows how many people want to run AI tools together on their own machines. But this is a stage where users report 8% of the usage limit vanishing per order entry and the conversation being re-read on one in four turns. If you try it, do not connect it to important company systems. Start by giving it one simple task and checking for yourself how much of your limit it used.