TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A community report says Claude Code periodically sends background-agent transcripts to generate short status-line summaries, and measured those requests at 16.4% of its estimated API spend during a week of observation. Its author published a Bun preload that suppresses the timer; the measurements and reported savings are not independently verified here, and subscription users may experience the impact as rate-limit headroom rather than a lower invoice.
Flamerix AI has published a community-made Bun preload intended to suppress Claude Code’s recurring requests for short background-agent status summaries, after its author reported that the requests accounted for 16.4% of estimated Claude Code spend during a measured week. The reported cost is not an independently verified figure or necessarily an added subscription charge: the report says subscription users lose rate-limit headroom, while the proposed hook is an unofficial workaround rather than an Anthropic setting.
The report says Claude Code runs a summary task about every 30 seconds for each background agent. It describes the task as sending the agent’s full transcript to a model, usually Opus, and asking it to describe its most recent action in three to five words. The resulting text appears in the agents panel’s status line, according to the author, and is not passed into the user’s conversation.
Flamerix AI’s no-agent-summary.js is a roughly 70-line Bun preload that intercepts a timer callback identified by text associated with the summarizer. The author says the hook disables that timer while leaving other timers alone and does not modify the Claude Code binary. It is distributed under the MIT License. Because it changes runtime behavior through Bun’s preload mechanism, it is an unofficial workaround, not a supported product control.
In the report’s OpenTelemetry data, events labeled query_source = agent_summary represented 16.4% of estimated spend during the week before the hook was used and 19.7% of API requests. The author says the share fell to zero two days after enabling the hook. These are the report’s own measurements; its stated spend figure is Claude Code’s per-request estimate, not an invoice total. The report also points to an issue collecting four other measurements in the 18%–26% range, but those figures are not independently established by the supplied material.
Status Summaries Use Model Calls
If the reported pattern applies to other users, recurring summaries could take a meaningful share of a Claude Code user’s available API capacity while producing text that is only shown in the agents panel. That matters most for people running several background agents or working under usage limits: a request that does not advance the main conversation may still consume tokens and, for subscriptions, reduce the capacity available for other work.
The distinction between estimated spend and actual billing matters. The report does not show that every user will save 16.4% on a bill, and the figure is tied to one author’s telemetry and observation window. Nor does the workaround establish that disabling summaries is appropriate for all workflows; users may value the status updates even if they cost capacity.
Products Worth Considering
EcoFlow Smart Home Panel 2 Whole-Home Backup System Hub
Logitech MX Mechanical Wireless Illuminated Keyboard Tactile (Graphite)
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How the Report Measured Usage
The author says the issue was identified by examining Claude Code’s OpenTelemetry export and grouping API-request events by their query source. The report places the behavior’s default activation at version 2.1.232, which it dates to August 2026, and says its own testing covered versions 2.1.257 through 2.1.312 on Linux. Those version dates and test details are claims in the supplied report, not independently confirmed release history.
The preload relies on Bun’s ability to load a JavaScript file before the application. According to Flamerix AI, editing the bundled code is not a practical route because Claude Code ships as precompiled Bun bytecode, so the workaround instead identifies and stops the relevant timer. The author reports that others have seen it work on macOS, but says the hook’s behavior depends on how Claude Code is installed and must be checked after updates.
“Every ~30 seconds, for every background agent, Claude Code sends that agent’s entire transcript to the agent’s own model.”
— Flamerix AI, in the report
Scope and Savings Need Verification
The supplied material does not independently confirm the reported request pattern, the 16.4% spend share, or the reduction after installing the preload. The figures reflect one author’s telemetry and a particular measurement period; the report does not provide enough detail here to establish how usage, agent counts, or model choices affected the result. Its reference to other measurements is not a substitute for reviewing those measurements directly.
It is also unclear how broadly the hook works across operating systems and installation methods. The author says Linux was tested and macOS was reported working by others, while macOS and Windows installation shapes were not measured. An update may undo the workaround, and the report advises checking that it remains active after each update. The supplied source does not say whether Anthropic has confirmed the behavior or plans to add a setting to control it.
Check the Hook After Updates
For users evaluating the workaround, the report says to restart the Claude Code daemon and existing sessions, launch Claude Code with BUN_OPTIONS set, and verify that the preload has taken effect. It warns that the environment variable reaches Bun processes broadly, so installation choices and the exact preload syntax matter. Users should consult the original instructions and assess the implications before applying a community script.
The next useful evidence would be independent measurements across versions, operating systems, and different numbers of background agents, along with clarification from Anthropic on whether summaries can be disabled through a supported setting. Until then, the reported savings should be treated as a user-measured result, not a guaranteed reduction in token use for every Claude Code installation.
Key Questions
What is Claude Code reportedly spending tokens on?
Flamerix AI says Claude Code periodically sends each background agent’s transcript to a model to generate a short status summary for the agents panel. The claim has not been independently verified in the supplied material.
Does the 16.4% figure mean users pay 16.4% more?
No such conclusion is established. The report calls 16.4% a share of Claude Code’s estimated spend for a measured week and says the estimate is not an invoice. For subscription users, it describes the cost as reduced rate-limit headroom.
What does the community workaround do?
The Bun preload published by Flamerix AI is designed to stop the timer that triggers agent-summary requests. It is an unofficial script and does not modify the Claude Code binary, according to its author.
Has the workaround been tested on every platform?
No. The report says it was tested on Linux with Claude Code versions 2.1.257–2.1.312 and that macOS users had reported success. It says macOS and Windows installation configurations were not measured, so compatibility should not be assumed.
Source: gist.github.com
