CloudConvert

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build me a file conversion service that replaces CloudConvert: an API and a queue that runs the right tool for each format, sandboxed, on my own machine.

STACK
- Node 20+ with Fastify for the API
- SQLite through better-sqlite3 for jobs and history, WAL mode
- The converters are existing binaries, each in its own container: ffmpeg, ImageMagick, LibreOffice, Ghostscript, pandoc, qpdf, 7-zip
- Docker or Podman on the host; the API never runs a converter in its own process
- Caddy in front

THE DATA MODEL
- jobs: id, token, status, created_by, created_at, started_at, finished_at, error, priority
- tasks: id, job_id, kind, engine, input_file_id, output_file_id, options_json, status, attempts, log_path, started_at, finished_at
- files: id, job_id, kind, original_name, stored_path, mime_detected, mime_claimed, size_bytes, sha256, expires_at, deleted_at
- engines: name, image, formats_in_json, formats_out_json, cpu_limit, memory_limit, timeout_seconds
- usage: id, api_key_id, bytes_in, bytes_out, seconds_cpu, created_at
- A job is a graph of tasks: import, convert, optimise, export. Chaining is the reason this is a service and not a shell script

SANDBOXING, WHICH IS THE ACTUAL PRODUCT
- Every conversion runs in a container with no network, a read-only root filesystem, a tmpfs for work, a memory cap, a CPU quota, a wall-clock timeout and a dropped capability set
- The input is mounted read-only at a fixed path; the output directory is the only writable mount
- Non-root inside the container, and a seccomp profile
- File parsers are among the most reliably exploitable code in existence. Ghostscript, ImageMagick and LibreOffice have all been remote code execution in the last few years. The isolation is not caution, it is the requirement — write that in the README
- The detected MIME type governs which engine runs, never the file extension and never the client's claim; a mismatch is recorded

THE API
- POST a job describing tasks, or use the one-shot convert endpoint for the common case
- Upload directly, by URL, or with a pre-signed form; download by a signed link that expires
- Poll for status, or receive a webhook signed with HMAC and retried with backoff
- Errors say which task failed, with the last lines of the engine's log, sanitised of paths
- Idempotency keys, so a retried request never runs the job twice
- API keys with per-key limits on concurrency, file size and monthly bytes

WHAT IT CONVERTS
- Images: raster and vector both ways, with resize, crop, quality, colour profile handling and metadata stripping optional
- Documents: office formats and PDF through LibreOffice, PDF operations through qpdf and Ghostscript, markup through pandoc
- Audio and video through ffmpeg, with sensible presets rather than a hundred exposed flags
- Archives, extracted with a limit on the expansion ratio and the entry count — a zip bomb is a denial of service with a two-line fix
- Every format pair is declared in the engines table, so the API can answer 'what can you convert this to' from data rather than from a hard-coded list

THE QUEUE
- Workers pull tasks, with a concurrency cap per engine and a global one
- Retries with backoff for transient failures, never for a parse failure
- Priority so a small image is not stuck behind an hour of video
- Progress reported where the engine reports it, and honestly absent where it does not

FILES AND PRIVACY
- Everything expires: a default of an hour, configurable per job, enforced by a sweeper that also deletes on disk
- Deletion is real, and the row records that it happened
- Optional at-rest encryption with a per-job key held only for the job's life
- The README states plainly how long files live and what is logged

THE INTERFACE
- A page to drop a file, pick a target format, and get the result — the API with a face on it
- A job list with logs for the operator
- Dark and light

OPERATIONS
- .env: DATABASE_PATH, STORAGE_PATH, BASE_URL, SIGNING_SECRET, MAX_UPLOAD_BYTES, CONTAINER_RUNTIME
- Migrations on boot, each once
- Disk watchdog: refuse new jobs above a threshold rather than filling the volume
- Health endpoint that converts a one-pixel PNG through the real path
- Backup covers the database; the files are ephemeral by design and are not backed up

WHAT MATTERS MOST
The sandbox. Build the container runner with its limits and the MIME detection first, and try to break it with a malformed PDF and a zip bomb before writing a single format handler. Everything else here is calling a binary with the right arguments.

Give me the repository, the container definitions, migrations, .env.example, and a README with deploy steps behind Caddy and the security assumptions stated in full.