The problem
Rate limits on bulk API work are painful. You
start processing a spreadsheet row by row, hit
a 429 after a few requests, wait, continue,
hit it again. You end up babysitting an API
for hours.
What I built
Claude Batch is a paced request queue in a
single HTML file. No build step, no npm install,
no server. Open it in a browser and it works.
Two input modes:
- Type prompts (one line = one request)
- Upload CSV/Excel and write a {ColumnName} template — one request per row, only the columns you reference get sent (typically 70–80% fewer tokens)
Set an interval (1–180s), press Start, walk
away. Exports as JSON, CSV, or plain text.
Engineering details
- Retry with exponential backoff (2s, 4s, 8s, cap 20s)
- Respects retry-after header — server guidance overrides calculated backoff
- AbortController on every fetch — cancels the actual HTTP connection on Stop
- 504 in the retryable set (gateway timeouts are transient)
- sleepAbortable() polls isRunning every 100ms during backoff — Stop resolves immediately
- 29 tests across 95 assertions in a jsdom harness that loads the actual index.html and drives real UI events
10 real bugs found and fixed
Every bug has a regression test:
- No retry on 429s — fixed with backoff logic
- Timer leak on repeated start/stop — fixed with tracked setInterval
- Stuck processing row after Stop — fixed with status reset sweep in stopRun()
- Clear All orphaning in-flight rows — fixed by disabling during run
- retry-after header ignored — now read and respected
- No AbortController on fetch — added
- 504 missing from retryable set — added
- Draggable state not synced with isRunning
- stopRun() only re-rendered conditionally
- Dead code currentRowTimer — removed
Links
Live: virerra.github.io/Claude-Batch
Source: github.com/Virerra/Claude-Batch
License: MIT
Top comments (0)