agentburn

Hermes Agent · OpenClaw · Claude Code — one normalized core, local, read-only, zero dependencies

uvx agentburn

Why this exists

Always-on agents bill you around the clock — and their built-in counters only show totals. Real threads that made this tool:

"73% of every API call is fixed overhead — ~13.9K tokens of tool definitions and system prompt, resent every time." — hermes-agent #4379

"One entrant wrote about waking up to a $47 surprise bill from an overnight run — that's not an exotic failure, it's the default behavior of an unsupervised loop." — dev.to

"I've seen runs where step 3 costs 4× step 1 — no alert, just a bill." — comment, ibid.

agentburn reads the agent's own accounting data (read-only) and answers the question the totals never do: where.

Related MCP server: openclaw-cost-tracker-mcp

What it answers

• Where it burns — by source: cron / subagent / gateway:telegram|discord|whatsapp / cli. Always-on ≠ free: scheduled jobs and gateways spend without you.

• 🌙 While you slept — the overnight bill, isolated and named (configurable window: --night 23-7).

• Fixed overhead — average input tokens per API call per source. The "73% overhead" pattern is visible in one glance; with request dumps enabled, you get the sampled composition (system prompt vs tool definitions vs history).

• Subagent rollups — delegation cost chained back to the session that spawned it. Recursion compounds; here is the receipt.

• Top tools — which tool results weigh most in your context.

• What to do — up to 4 conservative, named recommendations with monthly estimates.

How it compares

As of June 2026; ccusage and codeburn are excellent at what they do — agentburn deliberately starts where they stop (ccusage scoped per-tool analysis out).

Why trust these numbers

Most token trackers quietly disagree with each other (2–91× in public issue threads). agentburn takes the opposite stance:

• Numbers come from the agent's own accounting (~/.hermes/state.db: per-session token counters and cost fields). No scraping, no proxies, no guessing.

• Provider-billed costs are shown as-is; Hermes estimates are marked with ~. Mixed data is labeled mixed.

• Sessions with messages but zero recorded tokens (known Hermes accounting gaps, e.g. #12023) are detected and reported: totals are then explicitly a lower bound — and fixing the accounting becomes recommendation #1.

• Input composition from request dumps is char-proportional and labeled sampled estimate, not truth.

Privacy

Everything runs locally and reads your database read-only. No network calls. No telemetry. The report is yours.

Usage

agentburn                        # every agent on this machine, last 30 days
agentburn --agent openclaw       # just one
agentburn --days 7
agentburn --agent hermes --db /path/to/state.db
agentburn why                    # behavioral forensics: loops, retry storms, idle heartbeats
agentburn why --source telegram  # decompose ONE source: functions called, errors, loops
agentburn --source cron          # cost report for one source only
agentburn explain --model llama3.1   # LLM reads the numbers back to you (local by default)
agentburn --night 23-7           # custom overnight window (local time)
agentburn --budget-month 50 --fail-over   # sentinel for cron/CI
agentburn --json                 # machine-readable, pipe it anywhere
agentburn --no-color

Mechanics

📤 Share your burn (--share). An anonymized card — categories, models and totals only; session titles, paths and content are excluded by construction. Safe to paste into a post; --svg card.svg renders the same card as an image:

🔥 my hermes agent · last 30d
~$45.50 → ~$430/mo pace · 1.75M tokens
where it burns: cron 79% · cli 9% · telegram 7% · subagent 5%
🌙 while I slept (00–08): ~$36.00 — 79% of everything
⚙️ telegram re-sends 20,000 tokens with EVERY call — 2.5× the community norm (≈8k)
— agentburn · local & private

--svg card.svg renders it as an image:

📏 Calibration against public benchmarks. "Is 15k input tokens per call normal?" The report compares your fixed overhead with community-measured references embedded as dated constants (e.g. the Phala always-on-agent benchmark, 2026-03: ≈8k/call baseline). No network — sources are cited inline.

📐 Optimize → prove it (--save-baseline / --compare). Snapshot your pace, change the config (cheaper cron model, trimmed toolsets), then agentburn --compare shows the delta in $/month — pace-normalized, so a 7-day baseline compares honestly with a 30-day window. Every recommendation becomes a testable promise.

🔬 agentburn why — behavioral forensics. report says where it burns; why says why, from the agent's own recorded actions and thoughts:

🔬 agentburn why — openclaw · gateway:telegram

   WHAT IT ACTUALLY DID   browser 34× ≈210K in results · web_search 18× · shell 7× (2 errors)
   RE-READ LOOPS          5× browser(https://news.site/page) — every repeat re-paid in full
   RETRY STORMS           Bash: 3 errors / 6 calls — paying full price for every error
   IDLE HEARTBEATS        4 of 9 heartbeat runs did NOTHING — $2.40 of pure idle burn
   BURNED ON FAILURES     2 failed runs → ~$3.90 (timeout, killed)
   THINKS MORE THAN IT WORKS   62% thinking · 84K tokens · "rename files task"

   💡 WHAT TO CHANGE
   1. `/proj/big.md` was fetched 4× in one session ≈32K tokens re-paid — cache it…

Observations with numbers, not verdicts; only tool names, truncated argument keys and counters — message content never leaves the machine (and never enters the report).

🧠 agentburn explain — LLM interpretation, local-first. The numbers, read back to you in plain language with ranked actions:

agentburn explain --model llama3.1                      # local ollama — nothing leaves the machine
agentburn explain --llm https://openrouter.ai/api/v1 \
  --model deepseek/deepseek-chat --yes-remote --lang ru # remote: explicit opt-in only

Privacy rules are hard-coded: the default endpoint is localhost (ollama / LM Studio); a remote endpoint requires --yes-remote and receives a redacted summary only — session titles become session-N, file paths shrink to basenames, message content is never in the payload to begin with. Works with any OpenAI-compatible API, zero new dependencies. (Yes — a cost profiler spending ~3K tokens to explain costs. The payload is compact and the answer capped; the irony is acknowledged.)

🔧 agentburn fix — from findings to ready config patches (dry-run by design). Not "consider a cheaper model" but the exact file and the exact lines:

🔧 agentburn fix — hermes · DRY-RUN (nothing was changed)

   1. Point Hermes cron jobs at a cheap model
      file   : ~/.hermes/cron/jobs.json
      why    : cron is 79% of spend; maintenance rarely needs a frontier model.
      effect : bulk of ≈$341/mo moves to cheap-model pricing
      proposed:
        "nightly digest": "model": "deepseek/deepseek-chat"
      ⓘ field verified in hermes-agent cron/jobs.py: per-job `model` override

Patch generators exist only for config keys verified against the agents' source code (Hermes cron/jobs.json, OpenClaw agents.defaults.heartbeat incl. activeHours — the night-burn killer — and lightContext). There is no --apply on purpose: paste it yourself, then prove the saving with --save-baseline → --compare.

🔌 agentburn mcp — your agent answers for its own bill. A zero-dependency MCP stdio server exposing burn_report / burn_why / burn_card. Register it and ask the agent "where do you burn my money?" — it calls the profiler on its own database and explains:

# Claude Code
claude mcp add agentburn -- agentburn mcp
# Hermes / OpenClaw: add an stdio MCP server with command `agentburn mcp`

🩺 agentburn doctor. Trackers disagree because the agent's own accounting has gaps. doctor names the broken combinations (provider × model × source) for zero-usage and unpriced sessions, and generates a ready-to-paste upstream bug report — counters only, no message content.

🚨 Sentinel mode — a budget guard for server agents. Your agent runs 24/7 on a VPS; this watches it:

# alert when overnight burn exceeds $5/month pace (exit code 1 → any alerting hooks in)
agentburn --agent openclaw --budget-night 5 --fail-over --no-color \
  || notify-send "🚨 agent is burning money at night"

Drop it in cron next to the agent itself — the one-off check becomes a standing guard.

Supported agents

One normalized model, one adapter per agent. Run agentburn and every agent found on the machine gets its own report.

Adapters are ~150 lines over a shared model. Codex CLI / opencode are natural next targets — PRs welcome.

Related

token-history — the macro view: daily archive of which agents the world uses (OpenRouter rankings). agentburn is the micro view: where yours burns.

License

MIT

the token-* family · token-history — which agents the world runs · agentburn — where yours burns

if this saved you a dinner's worth of tokens, a ⭐ helps the next person find it