Orbit — one workspace for every cloud drive you own

Loading project

A multi-cloud file manager that puts Google Drive, OneDrive, Dropbox, pCloud, MEGA and any S3-compatible bucket behind one interface, so files can be browsed, searched, moved and shared across all of them at once. Orbit stores no file bytes of its own — only metadata and encrypted tokens — and streams each file from the provider it already lives in. Live, and free to run.

Orbit aggregates every cloud drive a person has accumulated into one workspace, without becoming another place their files are stored.

What it does. Connect several accounts — including several of the same provider — and browse them through one provider-agnostic path. Search reaches across every connected drive rather than the page in front of you, filtered by type, size, and age bands for both modified and created date. Files move and copy between providers, not just within one. Bulk actions can select a page or every page, and run as background jobs that keep going, and keep reporting from the header, when you navigate somewhere else.

Beyond browsing. A storage breakdown of what is actually using the space — photos, video, audio, documents, archives, code — per account and overall. A cross-cloud duplicate finder. Collections, which are virtual folders spanning drives. Short share links on your own domain with QR codes, where the provider's URL is never exposed. Scheduled jobs. A bin merged across every drive that keeps one. Browser-to-browser direct transfer over WebRTC. And a developer platform: a versioned public API, OAuth apps with PKCE, webhooks, and in-app documentation generated from the same Zod schemas the routes validate with.

What I learned

A pipeline's exit code belongs to the last command. I checked lint with `npm run lint 2>&1 | tail -8; echo $?` and read zero — which is tail's status, not npm's. Five real errors went in, and "0 errors" got written into the repo's own verification table before CI disagreed.

drizzle-kit reads .env from its own working directory. That is packages/db, which has none, so DATABASE_URL came back undefined, the config fell back to the local SQLite file, and `npm run db:migrate` printed "migrations applied successfully" while production stayed a version behind. Identical output either way. The habit that caught it is now written down: query the schema afterwards, never trust the message.

I read a database as something it was not. The metadata mirror had existed since the first sync engine, so I pointed folder browsing at it — a large speed win on paper. It has no folders: every adapter filters them out of its flat enumeration, because the mirror was built for the storage breakdown and the duplicate finder, which need names and sizes and never a path. Worse, Drive files were all filed at /{name}. In production that was 4,860 of 5,214 rows sitting at the root. Browsing showed a flattened drive, and files appeared to come back every fifteen minutes as each sync pass rewrote their rows. Pressing Refresh went to the provider and looked correct, which made a data problem read as a display glitch. One SELECT would have shown it in seconds and I never ran it.

Google answers a rate limit with 403 about as often as 429 — userRateLimitExceeded — and 403 is also how it says "this file is not yours". Treated as one thing, a throttled file in a bulk delete was reported as permanently undeletable. They differ only by a reason string in the body, which has to be read once and reused, because a Response body cannot be consumed twice.

A proxy timeout looks like a hang, not a failure. Cloudflare sits in front of the API and closes a long request; its documented limit is 100 seconds. Deleting 167 files one at a time, at roughly a third of a second each, ran past it — so the browser waited on a request that had already been cut, the progress bar sat at zero, and nothing was ever reported either way.

A test can pass locally and be incapable of passing on CI. One counted provider cards on a page. My .env has Dropbox keys, so Dropbox appeared once; CI has none, so it also appeared in the "not set up yet" list and appeared twice. Reproduced by unsetting the keys locally, which is now the first thing to try when CI and a local run disagree.

Serial is not a safe default. The E2E job ran 68 minutes and got cancelled twice by its own timeout, because CI was pinned to one worker while the same sweep took four and a half minutes locally in parallel. Twenty-five of those specs were also reaching for a sidebar that a phone does not have, at 60 seconds a time, three attempts each.

An XML comment cannot contain two hyphens in a row. Naming CSS custom properties inside an SVG comment made the file invalid XML. Browsers render it happily and strict parsers refuse it, which is exactly how a broken image reaches a README unnoticed.

Orbit — one workspace for every cloud drive you own

The problem

What I built

Built with

What I learned

Gallery