See why your AI agents fail, and what they cost.
Earlyman reads your agent traces on your Mac, names the step that broke the run, and shows what your agents and models cost.
Free and open source. No account. Runs on your Mac.
- Install
- Open your dashboard
- Ask from Claude Code
Interactive demo of the Earlyman dashboard
Good morning. 4 things need you. Most urgent: a stuck agent is still costing money.
- An agent has made no progress for 2 hA tool call is still pending. Stop it or start it again.$3.10 / h
- PR #214 broke a rule: force push to mainBlocked before it ran. The agent carried on.1 rule
- Search subagents run on the largest model418 of 483 this week. A smaller model is enough for search.−$38 / week
- This session has run for 9 hHand off to a fresh session to stop paying for old context.−$6 / day
All caught up.
-
Tokens2.4MSessions2ModelOpusRuleforce push to main · blocked
-
Tokens910kSessions1ModelSonnetRulesall kept
-
Tokens7.1MSessions7ModelOpusRulesall kept
-
Tokens380kSessions1ModelHaikuRulesall kept
Click a pull request to open its receipt.
- checkout-retryOpusstuck 2 h$3.10 / h
- invoice-exportSonnetrunning$1.20 / h
- search-docsOpusrunning$2.40 / h
- test-fixerHaikurunning$0.18 / h
- 8 moremixeddone$0.00 / h
- No force push to mainblocked PR #2141 hit
- No secrets in commits0 hits
- Ask before deleting files3 hits
- Stop an agent idle for 30 m2 hits
3 of 4 rules on.
Last 7 days · $126 · busiest Thursday, $31
- Opus54 %$71.40
- Sonnet31 %$18.90
- Haiku11 %$1.30
- Local 8B4 %$0.00
A smaller model is enough for 418 search calls on Opus this week. Move them to Haiku: −$38 / week
- Claude Code
- Codex
- Anthropic API
- OpenAI API
- OpenRouter
- RunPod
- Ollama
- LM Studio
- vLLM
- llama.cpp
- Runs on your MacNo account and no Earlyman server
- Counts, not contentNever stores or sends your prompts or replies
- Agents, APIs and modelsClaude Code, cloud bills, GPU pods and local servers
- Free and open sourceApache-2.0, no meter
From one agent step to the monthly bill.
- 01Your agent works.Earlyman counts every step, tool call and token on your Mac, as it happens.
- 02The pull request gets a receipt.What it cost, how long it took and which rules it kept, next to the code you review.
- 03The month stays on budget.Receipts add up to spend and limits across agents and vendors, with a heads-up before the wall.
See a pull request get built
An agent works, the pull request gets its receipt, and the month stays on budget.
- Plan the changecostliest41 s$1.90
- Read the code6 s$0.08
- Run the tests failedslowest120 s$0.31
- Edit six files38 s$1.40
- Run the tests again passed22 s$0.12
The first test run failed and sat on a two-minute wait. That one step took 53 % of the session.
Run only the failing test first, with a 30-second limit. One line for your agent's instructions, so the next run does not wait.
The same failure showed up in 17 of 57 runs this month. One fix covers all of them.
When steps are missing from a trace, Earlyman says so and lowers how sure it is about the cause.
Technical details
Step 3 of 5 · run tests · exit code 1 · 120 s · 1 failed tool call, recovered on the next step. Planning cost the most: $1.90 of $3.81.
Open any pull request and see only its traces. Works on Claude Code sessions, SWE-agent and OpenHands runs, and OpenTelemetry or Traceloop spans. How it finds the cause →
It names the cause of a failed run. We checked on runs it never saw.
- Cause named right
- 88 %
- Good runs shown as failed
- 5 %
- Set-up needed
- 0
35 of 40 failed runs, judged blind against each run's own tests and patch. The other 5 were partly right; none were wrong.
2 of 43 runs that solved their task.
No signals or labels to write first. The runs were imported as they were.
The runs: public coding-agent runs from SWE-agent and OpenHands (Nebius, CC BY 4.0) and SWE-smith (MIT), held out from all tuning. When only the agent's own checks pass, Earlyman marks the run unverified rather than fine. A run whose repo tests pass can still miss the task's hidden tests and look fine: that is the honest limit of reading a run without them.
Side by side: on the same 40 failed runs, blind judges preferred Earlyman's cause and fix on 39; the other was a tie.1
Held-out set, scored 25 and 26 October 2026.
1 Compared with Raindrop's free plan with its signals turned on. Raindrop's paid triage was not tested.
- No force-push
- Ask before pushing
- No web fetches
We run our own agents on it. Here is the last month.
- Agent runs traced
- 1,259
- Model calls counted
- 96,651
- Input read from cache
- 96.5 %
- Changes put to the test
- 17
Each one with its steps, time and cost.
Across 49 working sessions.
Measured, so we can see when it drops.
2 rolled back. The rest still wait for their verdict.
From Earlyman on our team's own machines, 24 September to 25 October 2026. Counts only; spend left out on purpose.
Before you install it.
Does Earlyman store or send my prompts?
Only when you open a step to see its input and output, and only on your Mac: it never stores or sends them. Everything else comes from usage numbers, timings and tool names in the logs Claude Code already keeps. The step text is read from the local log on click and is not kept afterwards.
Is it free?
Yes. Earlyman is free and open source under the Apache-2.0 licence, with no account, no key and no meter.
Which agents does it work with?
Claude Code sessions, SWE-agent and OpenHands runs, and any agent that sends OpenTelemetry or Traceloop spans.
Where does my data go?
Nowhere. Earlyman runs on your computer and there is no Earlyman server. Billing checks and alerts only run once you set them up.
How is it different from cloud trace tools?
Cloud tools keep the text of your sessions on their servers. Earlyman keeps counts on your computer, and knows your plan's 5-hour and weekly limits. See the full comparison.
Earlyman never stores or sends your prompts.
It runs on your Mac and keeps counts, not content.