Features

One agent session,
start to finish.

Watch one sample session, recorded from the app. Each part of Earlyman shows up when it matters. Never your prompts.

Get started How it compares →
Recorded from the Earlyman app · sample data1 min 15 s · voice-over, captions on

Also in each step

Step 1

Usage & limits →

  • Gauges for your 5-hour and weekly windows, how fast you burn them, and when they reset.
  • Spend guard warns past one amount and stops the session at another.
  • Organisation API limits: how much room each key and model has left, and API spend against a monthly limit. opt-in

Step 2

Rules →

  • Your rules live in one file, checked as each tool call runs; a hard rule blocks the step.
  • Lint the file against a baseline rule pack before it runs.
  • In Claude Code, once you switch the check on, each tool call is checked as it runs: report mode logs it, enforce mode blocks it.
  • Compare what the file declared with what was observed, on each pull request in Claude Code.
  • For Codex, the matching instructions, approval settings and command rules, written to a folder you pick.

Step 2

Traces →

  • Explain any run, from rules on its numbers.
  • Find the cause on any agent run: Claude Code sessions, SWE-agent and OpenHands runs, and OpenTelemetry or Traceloop spans.
  • Failures across runs: which tools fail, and how, by project, agent and model.
  • Incomplete-trace warning: when steps are missing, it says so and is less sure of the cause.
  • Fix for next time: one line for your agent's instructions, from the failure it prevents.
  • Claude Code and Claude SDK traces from the trace export built into Claude Code or the Claude SDK. opt-in
  • Ask in plain English, with a local model. off by default

Step 3

Hand-off · Beta →

Beta: summaries can miss details; check the transcript link.

  • Four short lines: the goal, what is done, what is next and what to avoid.
  • Carry on anywhere: a fresh Claude Code session, another Claude model or another tool such as Codex.

Step 4

PR receipts →

  • A receipt on the pull request: cost, tokens, agent time, failed calls and issues fixed with proof.
  • Cost per PR and per branch in the dashboard, with agent time and failures.
  • Kept current: the comment is edited in place as the branch moves, with a last update when it merges.

Step 5

Savings →

  • Experiments: a change is tried on part of your sessions and counts only when the difference is clear.
  • This week against last, each number marked measured or estimated.
  • Canary a change on a share of sessions first.

Step 5

Model registry →

  • The models the work ran on, and what each cost, next to the ones you fine-tune.
  • Your fine-tunes: versions, test scores, training spend and where each one runs.

Step 5

Spend and providers

  • Every bill on one screen: Claude Code next to API bills, plans you subscribe to, GPU pods and local models.
  • Plans found for you: services spotted from your key names and connected tools, each with its price, quota and renewal. Accept or dismiss.
  • History you keep after Claude Code clears its logs. Export any time.
  • Team totals: Self-hosted, off until you turn it on. Totals and metadata only; it will not start with content mode on. opt-in

Step 6

Failures and reliability →

  • Success rate per agent and model; a repeat failure becomes one issue, with its fix. A catalogue of 51 failure classes names what went wrong.
  • Issues with how often they happen and the trend.
  • Similar failures across projects, matched on what failed, never on your text. Local embeddings are opt-in
  • Keep a fix fixed: one click turns an issue into a check on later runs.
  • Detectors that keep score from your Right and Wrong marks.
  • Local judge on your Mac, unverified. opt-in
  • Behaviour signals: stuck loops and re-asks, without reading a word. partial

Step 6

Your agents →

  • Traces from the agents you ship, over OpenTelemetry. Prompt, reply and tool-argument fields are dropped before anything is written. off by default
  • Prompt-blind at the door for every trace that arrives.
  • Agent self-reports: a broken tool or missing permission lands in Do now.
  • User feedback on each run; a comment is kept only as its length.
  • Python and TypeScript add-ons that read usage numbers only. coming soon

Step 6

Alerts and access →

  • Alerts to Slack, a signed webhook or a script; off until you add a channel.
  • Ask from Claude Code about a failed run without leaving your session.
  • Read-only query API: sessions, verdicts, failures and fix lines for your scripts and dashboards. One token you can revoke; never step text.

Step 6

Privacy →

  • Runs on your Mac, with no Earlyman account or server.
  • Counts, not content: step text is never stored.
  • How it compares side by side with other tools, with sources, below.

How it compares.

Where Earlyman leads, and where it does not yet. The short version →

Earlyman compared with Langfuse, Laminar, CodeBurn, ccusage, Datadog
LangfuseLaminarCodeBurnccusageDatadogEarlyman
Plan limits, with runway before you hit them
Repeat failures grouped across runs?
Every saving proven on a verdict date
GPU pods, API bills and local models in one place
Works without storing your prompts
Alerts pushed to Slack or a webhook?

Hover a mark for the claim and its source. Sources read 2026-10-20: Langfuse · Laminar · CodeBurn · ccusage · Datadog.Vendor claims are not independently tested.

Compare all, with notes and sources
Earlyman compared with Langfuse, Laminar, CodeBurn, ccusage, Datadog
LangfuseLLM observability, cloud or self-hostLaminaragent observability, cloudCodeBurncoding-agent cost, CLI + desktopccusageusage CLIDatadogLLM observabilityEarlymanlocal console, verified in code
Agent observability
Per-run trace with tool calls and sub-agentsYSDK / OTel traces; its Claude Code plugin captures every prompt and messageYTranscript, tree and timeline; sub-agents as cardsPCost down to session and turns; no span treeNUsage tables onlyYProduct page: trace coding agents on your machine; details not verifiedYSpan tree per session, sub-agent cards with duration, tokens and $; sizes and labels only; your own agents too, over OpenTelemetry (off by default, content dropped on arrival)
Live and stuck agentsNNot listedYBrowser session recordingYMenu bar view, always in sightPLive 5-hour block monitor?Not statedYLive / quiet / stuck from file tails every 10 s; stuck rows carry a fix line
Cost on each step of the traceYCost per observationYCost badge per step, cache tiers pricedPPer turnNNot listed?Not statedYPer-step $ chip that sums to the session $
Cost and limits
Subscription 5-hour and weekly windows, with runwayNNot listedNNot listedYPlan, month pace, 5-hour and weekly windowsP5-hour billing blocksNNot listedYThree gauges with source and age, runway, reset time and model-switch advice
Cost by project, git branch, taskYBy metadataYBy user, session, customerYProject, model, git branch, task; yield per PRPPer project, with an instances option?Not statedPModel, project, agent type and git branch; no task label or cost per merged PR yet
Spend guard that stops a sessionNNot listedPUsage limits on their own billYguard: warn and stop hooksNNot listedNNot listedYA hook warns, then refuses further tool calls with the reason; one step removes it
History past the agent's own log cleanup, with exportY30 or 90 days by plan; APIY7 days to 6 months by plan?Not statedNReads the current logs?Not statedYLocal warehouse kept as long as your disk; CSV, JSON and OTLP export
Failures
First failure and one fix line per runNNot listedPFailure detector calloutNNot listedNNot listed?Not statedYFirst failure, the span path to it and one catalogue fix line, plus the three steps most likely to have caused the failure, ranked from step statistics with reasons
Recurring failures grouped across runsNNot listedYClustersPoptimize lists waste findingsNNot listed?Not statedYFingerprint groups with seen count, first seen, severity and a regressed flag; resolve, ignore, duplicate, and verified automatically when it stops recurring; similar failures across projects and clusters from a content-free signature (local embeddings opt-in); one click turns an issue into a check judged on every later run
Agent reports its own statusNNot listedNNot listedNNot listedNNot listedNNot listedYThe agent reports its own failure (broken tool, missing permission, capability gap) through the MCP tool; it lands on the run, in Do now and in alerts. The kind, tool and severity are kept; a note only as its length
Quality
Behaviour signals on every runYLLM-judge scoresYSignalsNNot listedNNot listed?Not statedPNine structural detectors, no classifier reading text; API only today. Opt-in, local content mode only: a local judge model on your Mac scores five meaning-based signals, labelled unverified
Evals and datasetsYEvals, datasets, annotation queuesYEval cases, labelling queuesNNot listedNNot listedYExperimentsPNot for agents, by design (needs prompt text). For your own models: eval scores with dates, current vs candidate
Success rate per agent and workflow, with a confidence intervalPScores per experiment run on a dataset, LLM-judge basedPLeads with reliable agents; eval runs compared per scoreNNot listedNNot listedPEvaluator dashboards on experiments; per-agent rate not statedYSix rates per agent type and model, workflow, project and branch, with a 95 % Wilson interval; counts only, no combined score
Alerts and search
Push alerts to a channelYThreshold alertsYSlack, on by defaultPDesktop trayNNot listed?Not stated on the page readYSigned webhook, Slack and a local command, plus macOS notifications and a daily digest; each alert carries counts, the fix line and a paste block, never text
Search across tracesYFilters and full textYSQL editorYEvery number drills downNNot listed?Not statedPSearch over titles, tool names, agent labels and failure classes; no numeric filters yet
Proof of change
Cohort A/B with a confidence intervalYDataset experimentsYRelease comparisonsPCompare periodsNNot listedYExperimentsYExperiments: any two cohorts of logged traces (model, service, branch, version, dates, before / after a change), 95 % CI, raw and Holm-corrected p-values, n per side; Coach and Playbooks holdout arms too
A verdict date on every applied changeNNot listedNNot listedNNot listedNNot listedNNot listedYBaseline, verdict date, kept or reverted; saved = 0 until a verdict lands
Team and breadth
Team roll-upYOrgs and rolesYWorkspacesNTeams is coming soon (waitlist)NNot listedYOrganisation accountsYSelf-hosted and opt-in: per-person and per-service totals and metadata only, no content
Coding agents readYClaude Code, Cursor, Codex and many SDKsYClaude Agent SDK, Vercel AI SDKY42 toolsY18 sourcesYClaude Code and CursorPClaude Code only today; Codex and Gemini CLI on the roadmap
Ask from the agent (MCP, CLI)YMCP and CLIYCLI and MCP with SQLYMCP serverNNot listedYMCP onboardingYMCP server, slash skills, status line and a status command with exit codes
GPU pods, API bills and local model serversNNot listedNNot listedNNot listedNNot listedYPart of the wider Datadog platformYRunPod pods with idle $, OpenAI / Anthropic / OpenRouter usage, local servers by port
Registry of the models you fine-tuneNNot listedNNot listedNNot listedNNot listedNNot listedYVersions, file headers, eval scores with dates, training spend, what is serving
Privacy
Stores your prompt textYYes, with maskingYYes, with redaction modesNNoNNoYYes, in your Datadog accountNNo. Optional local content mode, off by default, keeps short summaries only on your Mac and never sends them; a test plants a secret in a session log and asserts it never appears
Sends telemetry to the vendorYCloud, or self-hostYCloud; privacy mode off by default on lower plansPDesktop app sends an anonymous report, default on outside the EUNNoYCloud productNNo code path to us; outbound only with keys you add

Sources read 2026-10-20 · vendor claims not independently tested

Coming soon

Knowledge searchBehaviour pageCodex and Gemini CLI sessions

Install your way.

Free, no account, no key. Pick whichever you already use.

Homebrew

One command, updates with the rest of your tools.

coming with the first public release

uv

Run it once to try it, or keep it as a tool.

coming with the first public release

Claude Code plugin

Add it from inside Claude Code; the dashboard starts with your session.

coming with the first public release
Works with
  • 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
  • FreeNo account, no meter

Public install links arrive with the first release.

Our own team's usage

We run our own agents on it. Here is the last month.

Agent runs traced
1,259
Each one with its steps, time and cost.
Model calls counted
96,651
Across 49 working sessions.
Input read from cache
96.5 %
Measured, so we can see when it drops.
Changes put to the test
17
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.

A few more details.

Does Earlyman read my prompts?

No. It reads usage numbers, timings and tool names from the logs Claude Code already keeps. Everything it reads, and every call it can make.

Where does my data live?

On your Mac. There is no Earlyman account and no Earlyman-run server; the optional team server runs on a machine your team chooses. The only outside calls are to the billing services you give it keys for, and to an optional analyst that is off by default.

Which coding agents does it support?

Claude Code today, plus cloud API bills, GPU pods and local model servers. It also reads SWE-agent and OpenHands runs, and your own agents can send traces over OpenTelemetry or Traceloop. Codex and Gemini CLI are on the coming-soon list above.

Can my team use it?

Yes, two ways. Each person can run their own copy and let teammates open it over a private network with a passcode. Or, opt-in, one person runs a self-hosted team server: everyone gets their own key and a role, and the team sees totals and metadata only, per person and per service. No prompts, replies or other content are sent to it or stored.

What does it cost?

Nothing. It is free, with no account and no meter.

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.

See what your agents
really cost.

Get started Explore live demo →