Resource name: agentburn — https://github.com/Socialpranker/agentburn
Author: Socialpranker (disclosure: that's me)
Brief description: Local, read-only profiler that shows where a Hermes instance burns money. Reads ~/.hermes/state.db, splits spend by source (cron vs each gateway vs subagents vs CLI), isolates the overnight window, and goes below totals: why walks the message log for re-read loops, error retries and idle heartbeat runs; fix prints paste-ready config patches (verified against Hermes source — cron/jobs.json keys) without editing anything; doctor detects Hermes' own zero-usage accounting gaps (#12023-style) and turns totals into explicit lower bounds. Python stdlib only, uvx agentburn, MIT. Same normalized core also reads OpenClaw and Claude Code.
Category: Tools & Utilities
Why it's awesome: The Hermes tracker documents the pain this answers: the 73% fixed-overhead measurement in hermes#4379 and overnight-bill surprises. Built-in /usage reports totals; agentburn answers where and why, then proposes the config change with $/mo arithmetic from a real OpenRouter price snapshot. It's honest by construction: estimates are marked ~, accounting holes demote totals to lower bounds, nothing leaves the machine. Not a duplicate of agenttrace (already listed): agenttrace audits session health post-run; agentburn attributes money to sources/behavior and generates the fix. They compose.
Why now: Always-on setups (cron + gateways) are the default Hermes deployment in 2026, and that's exactly the configuration where one nightly job on a frontier model quietly becomes 79% of the bill. There's also a browser demo with fake data, so people can evaluate it without installing: https://socialpranker.github.io/agentburn/
Maturity label: your call — it's new (released this week) with a large offline test suite, so experimental or beta both seem fair.
Resource name: agentburn — https://github.com/Socialpranker/agentburn
Author: Socialpranker (disclosure: that's me)
Brief description: Local, read-only profiler that shows where a Hermes instance burns money. Reads
~/.hermes/state.db, splits spend by source (cron vs each gateway vs subagents vs CLI), isolates the overnight window, and goes below totals:whywalks the message log for re-read loops, error retries and idle heartbeat runs;fixprints paste-ready config patches (verified against Hermes source —cron/jobs.jsonkeys) without editing anything;doctordetects Hermes' own zero-usage accounting gaps (#12023-style) and turns totals into explicit lower bounds. Python stdlib only,uvx agentburn, MIT. Same normalized core also reads OpenClaw and Claude Code.Category: Tools & Utilities
Why it's awesome: The Hermes tracker documents the pain this answers: the 73% fixed-overhead measurement in hermes#4379 and overnight-bill surprises. Built-in
/usagereports totals; agentburn answers where and why, then proposes the config change with $/mo arithmetic from a real OpenRouter price snapshot. It's honest by construction: estimates are marked~, accounting holes demote totals to lower bounds, nothing leaves the machine. Not a duplicate of agenttrace (already listed): agenttrace audits session health post-run; agentburn attributes money to sources/behavior and generates the fix. They compose.Why now: Always-on setups (cron + gateways) are the default Hermes deployment in 2026, and that's exactly the configuration where one nightly job on a frontier model quietly becomes 79% of the bill. There's also a browser demo with fake data, so people can evaluate it without installing: https://socialpranker.github.io/agentburn/
Maturity label: your call — it's new (released this week) with a large offline test suite, so
experimentalorbetaboth seem fair.