Codenomics

Decode the true economics of your AI coding agents.

Every token dashboard tells you what you burned. Codenomics tells you what you got — the true cost of each shipped change, across Claude Code, Codex CLI, and Gemini CLI, computed entirely on your machine.

TRUE $/COMMIT = (compute $ + prompts × attention $ + active time × hourly $) / commits

A model that burns more tokens but needs fewer corrections and less babysitting wins on this number. That's the comparison that actually matters when you pick a model or an agent — and a token counter alone won't show it.

Quick start

npx codenomics init     # detect agents, write config
npx codenomics index    # scan local agent logs (incremental, seconds)
npx codenomics serve    # dashboard at http://127.0.0.1:3737

No account. No upload. Zero runtime dependencies. Your transcripts never leave your machine (PRIVACY.md).

What you get

• Local dashboard — true $/commit, API-equivalent burn, model-vs-model economics, daily burn chart, per-session drill-down with cache and context-bloat diagnostics, human vs machine (headless/CI) split.
• Multi-agent collectors — Claude Code transcripts, Codex CLI rollouts (per-turn token accounting validated against cumulative totals), Gemini CLI OTEL telemetry (best-effort; requires telemetry enabled).
• Economic drivers you control — what's a prompt of your attention worth? What's your loaded hourly rate? Set them in config or the dashboard; every metric updates instantly. Per-model pricing overrides included.
• Budgets — $ or token limits per day/week/month, globally or per project/agent. npx codenomics index --check-budgets exits nonzero on breach, so cron alerting is one line.
• Canned reports — npx codenomics report weekly|monthly renders Markdown + self-contained HTML with prior-period deltas, top sessions, budget status, and rule-based observations (e.g. "your headless jobs ran on a premium model; same tokens at haiku pricing = $X saved"). Optional Slack digest via webhook. npx codenomics report schedule --install-cron automates it.
• AI recaps (optional) — one-line session summaries via your own claude CLI, cached forever.

Commands

| command | what it does | |---|---| | init | detect agents, write user config; --from <dir> imports old claude-stats recaps | | index | incremental scan of agent logs; --check-budgets for cron | | serve | local dashboard (127.0.0.1 only by default) | | report weekly\|monthly | md/html artifacts; --slack posts a digest; schedule for cron | | config get\|set\|unset\|list\|path | dotted-path config editing | | summarize | AI recaps for recent sessions | | doctor | collector status, quarantined files, format-drift stats | | sync | preview of the future (opt-in, aggregates-only) team sync |

Configuration

~/.config/codenomics/config.json (user) and .codenomics.json (per-project), merged with flags > env > project > user > defaults:

{
  "drivers": { "attentionUsdPerPrompt": 5, "engHourlyRateUsd": 0 },
  "pricing": { "gpt-5.5": { "in": 1.25, "out": 10 } },
  "limits": [
    { "id": "daily-burn", "metric": "costUsd", "period": "day", "max": 50, "scope": "global" }
  ],
  "report": { "slackWebhookUrl": null },
  "server": { "port": 3737, "host": "127.0.0.1" }
}

$ figures are API-equivalent compute cost (cache-aware: read/write rates modeled per vendor). On a subscription plan your marginal $ is 0 — read them as a normalized compute meter that makes models and agents comparable. Built-in pricing drifts as vendors change it; override any model in config.

How it reads your data

| agent | source | notes | |---|---|---| | Claude Code | ~/.claude/projects/**/*.jsonl | full fidelity: cache 5m/1h split, commits, human/machine, prompts | | Codex CLI | ~/.codex/sessions/**/rollout-*.jsonl | per-turn tokens incl. reasoning; commits; subagent detection | | Gemini CLI | ~/.gemini OTEL telemetry | requires telemetry enabled; no commit/source signals (shown as —) |

These are undocumented, fast-moving formats. Parsers are tolerant by design: malformed lines are skipped, unknown event types are counted (doctor shows drift), and a file that fails to parse is quarantined without killing the run.

On first use, codenomics prints a one-time disclosure listing exactly which directories it scans (read-only), where it stores its own data, and that nothing is uploaded — then records your acknowledgement under the data dir. In non-interactive contexts (CI, service managers) it discloses and proceeds without prompting; set CODENOMICS_ACCEPT_DISCLOSURE=1 to accept up front and silence it.

Development

npm install
npm test        # tsc build + node:test (fixtures are scrubbed real transcripts)

Licensed FSL-1.1-MIT (Functional Source License): use it, read it, modify it, self-host it for anything except offering a competing product — and every release becomes plain MIT two years after it ships.

Contributions welcome — especially fixtures from agent versions we haven't seen. scripts/scrub-fixture.mjs redacts a transcript so you can share its shape without sharing its content.