Your coding agents, on a calendar.
See how well you really use your agents.
Every Claude Code and Codex session you ran on this machine, drawn hour by hour.
npx github:awlevin/codecal
npx github:awlevin/codecalIt serves http://localhost:4317 and opens it. No install, no build step, no config file, no account, no dependencies. Node 18 or newer is the only requirement.
You have a rough feeling about how you work with agents. Your transcripts have the receipts. codecal turns them into a scoreboard you did not have to keep.
How well do you parallelize? Every week shows two totals. Wall clock counts overlapping sessions once, so it is time you spent. Session time sums every block, so it is work that got done. The multiplier between them is how much agent work you ran per hour of your own, and peak parallel is the most sessions you had running at the same moment. Three agents at once on Tuesday afternoon is a number, not a vibe.
How long do you leave sessions hanging? Open any session and it tells you the median and the longest time you took to come back to it. That is the honest measure of context-switching: a session with a 4 minute median is one you drove, and one with a 3 hour median is one you abandoned and rediscovered.
Where does the time actually go? Blocks are colored by repo and worktree, so a glance shows which project ate the week, and whether you finished things or left five worktrees half-open.
Claude Code or Codex? Both are drawn on the same grid, with a toggle for either alone. The split is usually not what people guess.
And the ordinary questions too:
- What ran all night while I was asleep, and what did it cost?
- When did I really start, and when did I really stop?
- Which repo ate Tuesday?
- What was that session on Thursday afternoon, and how do I resume it?
Everything is read locally. Nothing is uploaded, and no transcript content leaves the machine.
An agent left running does not look like anything. It makes no noise, it opens no window, and the bill arrives at the end of the week.
On the calendar it is unmistakable: a stripe of identical little blocks marching down two or three days straight, through hours you were asleep. Click one and the panel names the damage, since it carries the models used, the assistant turn count, and the token totals for that session alone.
This is how I found a loop that ran unattended for two nights and quietly ate a week of credits. Nothing else I had would have shown me that, because every dashboard reports the total and none of them report the shape.
You took five minutes to reply to an idle agent. The session never stopped, but the transcript has a five minute hole in it. Draw those holes literally and one hour of work becomes two twenty minute blocks with a gap, and your week turns into confetti.
The idle gap slider is where you say how long a pause has to be before it counts as a break. Drag it and the calendar re-merges live.
| Idle gap | Blocks | Wall clock |
|---|---|---|
| 1 min | 314 | 13h 29m |
| 15 min (default) | 171 | 19h 47m |
| 120 min | 41 | 72h 02m |
Low values show individual turns. High values show "I had an agent open all evening". Fifteen minutes is a good middle, and the setting is remembered per browser.
| Week view | Seven days by hour, with a live now-line and a red marker on today |
| A block per stretch of work | Colored by repo and worktree, side by side when sessions overlap, widening into free space |
| Both agents | Claude Code solid, Codex striped, header toggles to show either alone |
| A scoreboard | Wall clock, session time, the multiplier between them, peak parallel agents, longest block, busiest day |
| Session detail | Span, time to return (median and longest), prompts, assistant turns, models, tokens, every block in the session, the transcript path, and a ready-to-paste resume command |
| Spot the waste | Runaway loops and sessions left running read as a stripe of identical blocks; the panel gives the models and tokens they burned |
| Hide the noise | A session that hung on an unanswered prompt for three days can be hidden in one click, and brought back with a toggle |
| Live refresh | The index rebuilds every 30 seconds, so a session you start while the page is open appears on its own |
| Control | What it does |
|---|---|
Today, ‹, ›, ←, →, T |
Move between weeks |
| Idle gap slider | Live re-merge of blocks |
| Subagents toggle | Count subagent runtime as session runtime |
| Claude / Codex toggles | Show one source or both |
| Hidden toggle | Bring hidden sessions back, drawn muted |
| Projects menu | Show or hide individual repos and worktrees |
| Click a block | Open the session detail panel |
Esc |
Close the detail panel |
npx github:awlevin/codecal --port 5000 # serve somewhere else
npx github:awlevin/codecal --no-open # do not open a browser
npx github:awlevin/codecal --no-codex # Claude Code only
npx github:awlevin/codecal --no-claude # Codex only
npx github:awlevin/codecal --no-cache # force a full re-read
npx github:awlevin/codecal --helpRun it from a clone instead:
git clone https://github.com/awlevin/codecal.git
cd codecal
node bin/codecal.mjs| Source | Read from | Becomes |
|---|---|---|
| Claude Code | ~/.claude/projects/**/<session>.jsonl |
An activity point per timestamped record |
| Claude Code | turn_duration records |
A real span, so one long tool call still reads as runtime |
| Claude Code | <session>/subagents/*.jsonl |
Subagent runtime, credited to the parent session |
| Codex | ~/.codex/sessions/** and ~/.codex/archived_sessions/** |
An activity point per rollout record |
| Codex | session_meta, turn_context, event_msg |
Working directory, model, prompts, token totals |
| Both | Session title, else the first typed prompt | The block label |
| Both | Working directory | The project color, worktrees included |
Points and spans are fused into intervals at index time, then merged into blocks in the browser at whatever idle gap you pick.
The largest records carry bytes rather than metadata, so they are read for their timestamp alone, and every file is cached in ~/.cache/codecal/ by size and mtime. A cold pass over ~2.5 GB of transcripts takes about ten seconds. After that a re-index is well under a second, since only changed files are re-read.
ChatGPT web conversations. Those live on OpenAI's servers, not on your disk, so there is nothing local to read. Only the Codex CLI and Codex Desktop write rollouts locally, and those are what this reads.
Transcript formats change between releases of both tools. If a week looks empty or wrong, the format probably moved. Open an issue with your Claude Code or Codex version.
demo/ holds the Remotion project that renders the video and the hero image at the top. Its sessions are synthetic, so no real work shows up in them.
cd demo
npm install
npx remotion studio # edit it live
npx remotion render Demo # out/codecal-demo.mp4MIT
