MCP server de bg0ne

bg0ne

BG, gone. Background removal that runs on open models, self-hosts in one container, and bills by the image instead of by the month.

Live at bg0ne.com. MIT licensed.

What it is

Background removal is a solved problem with excellent open models behind it, and it is still mostly sold as a subscription with credits that expire. bg0ne is the same capability with the pricing turned back into what it actually is: a few cents of compute.

• Free previews. Capped by pixels, not by a counter. No account, no card. A visitor can find out whether it works on their photo without meeting a signup wall.
• 3 cents a full-resolution image. Credits never expire. Nothing to cancel.
• Agents pay per call. POST /api/cutout answers 402 with an x402 offer and settles in USDC. No account, no key, no card.
• Or run it yourself. docker compose up. Leave the payment variables unset and it is simply a free private instance.

Models

Every model this ships is permissively licensed, because the hosted service is billed for and a non-commercial model would poison that.

Model	License	Role
u2net	Apache-2.0	preview tier, baked into the image
birefnet-general-lite	MIT	paid tier, baked into the image
birefnet-general	MIT	best quality, ~1GB, fetched on demand
birefnet-portrait	MIT	portraits, fetched on demand
isnet-general-use	Apache-2.0	available, fetched on demand

BRIA's RMBG is the obvious better-known choice and is deliberately absent. Its weights are CC BY-NC 4.0 and commercial use requires a paid agreement with BRIA. rembg will happily serve it under the name bria-rmbg, so the model allowlist in infer/server.py is what actually keeps it out of a billed service. That allowlist is load bearing, not decoration.

The full birefnet-general is better again but is roughly a gigabyte and wants a GPU to be worth the wait. Set INFER_HD_MODEL=birefnet-general once there is one.

Run it

git clone https://github.com/profullstack/bg0ne.com
cd bg0ne.com
docker compose up

That is the whole setup. The app migrates its own database on boot.

Without Docker you need Bun 1.3+, Python 3.12+ and a Postgres:

bun install
pip install -r infer/requirements.txt
DATABASE_URL=postgres://localhost/bg0ne bun run dev

Architecture

One image, two runtimes, talking over loopback.

browser / agent
      |
      v
  Bun + Hono  ......  Postgres (accounts, credit ledger, payments)
      |
      v  127.0.0.1:7001
  Python + rembg (ONNX)

The model is a separate process because every good open background-removal model is Python and ships as ONNX, and the web app is Bun. INFER_URL is what makes that a deployment decision rather than an architectural one: point it at a GPU box and nothing else changes.

Why the credit ledger is a ledger

credit_ledger is append-only and the balance is its sum. A single balance column is one bad UPDATE away from being wrong forever with no way to find out when it went wrong, and these are bought with real money, so people ask where they went.

Two consequences worth knowing before editing that code:

• A spend happens before the work, and takes select ... for update on the user row. Without the lock, twenty concurrent cutouts all read the same balance, all pass the check, and the account finishes negative having been given work nobody paid for. There is a test for exactly this.
• Because the spend comes first, a failed cutout has to be refunded explicitly. The refund is a fresh positive row rather than a deleted one: the ledger records what happened, and what happened is that somebody was charged and then made whole.

Credit grants are idempotent per payment, enforced by a unique index rather than by remembering. CoinPay retries a webhook until it is acknowledged, so "we already granted this" has to be a database constraint.

Credit codes and invites

• Credit codes (BG-XXXX-XXXX-XXXX-XXXX, from @profullstack/redeem-credits, stored only as hashes) add credits (one credit = one full-resolution cutout) to credit_ledger, once per account per code. Redeem on /account ("Have a code?") or POST /api/v1/credits/redeem {"code":"BG-..."}. Mint on the box: docker exec bg0necom-app-1 bun apps/web/src/admin/codes.js mint --credits 100 --label "for Ada" (list, revoke <id>, reconcile <user id>, quota <user id> <total>).
• Invites (@profullstack/affiliates): three single-use links per account at /invites, each /i/<code> for 30 days. The code rides the emailed sign-in link, and pays INVITE_BONUS_CREDITS (default 10) to both people only when that link creates the account, the two are not one mailbox, the new address is not a throwaway one, the inviter's account is a day old (INVITE_MIN_INVITER_AGE_HOURS) and the inviter has had fewer than 10 bonuses in 30 days.
• ?ref= links are first touch, kept in the referral_code cookie and recorded in aff_attributions when the link creates an account.

Environment

Variable	Required
DATABASE_URL	yes	the only hard requirement
SITE_URL	yes in prod	passkey rpID derives from it
ROLES	no	web,infer by default
INFER_URL	no	loopback by default; a remote URL stops the local model spawning
INFER_HD_MODEL	no	isnet-general-use
RESEND_API_KEY	no	without it, sign-in links are logged instead of sent
COINPAY_API_KEY	no	cp_live_/cp_test_; without it payments are off
COINPAY_BUSINESS_ID	no
COINPAY_WEBHOOK_SECRET	no
COINPAY_X402_KEY	no	a scoped key, not the merchant key above
X402_PAY_TO	no	the EVM address agent payments land in

Secrets belong in the vault and reach the container as platform variables. There is deliberately no .env loading in this app.

Tests

docker run -d --name pg -e POSTGRES_PASSWORD=test -e POSTGRES_DB=bg0ne -p 55433:5432 postgres:16-alpine
DATABASE_URL=postgres://postgres:test@127.0.0.1:55433/bg0ne bun test

License

MIT. The models carry their own licenses, listed above.