Arabic Pipeline MCP Server

Arabic Pipeline API

Pay-per-call utilities for Arabic text, audio and documents, sold over x402 in USDC on Base.

No API key. No account. No subscription. The first request returns 402 Payment Required; the client signs a payment and retries; the money lands in the provider's wallet. That is the whole protocol — and it is why this service can pay a provider directly, with no payment processor in between.

# Claude Code (any Streamable HTTP MCP client works the same way)
claude mcp add --scope user --transport http arabic-pipeline-mcp \
  https://arabic-pipeline-mcp.arabic-pipeline.workers.dev/

No install, no key: the tools are already on the other end. A tool call without payment answers with the x402 challenge, and your client pays it on Base.

Live endpoints

The MCP server exposes the same nine capabilities as tools, each behind its own x402 payment (text_metrics, text_normalize, text_detect_language, text_chunk, text_diff, text_search, audio_transcribe, doc_extract, doc_extract_batch).

Why not a generic text tool

Arabic breaks naive string handling in ways that are invisible until they cost you a failed match:

• Tashkeel — فَاطِمَة and فاطمة are the same name, but not the same bytes.
• Letter variants — أ إ آ → ا, ى → ي, ة → ه are folded on purpose, not by accident.
• Arabic-Indic digits — ١٢٣ must match 123, and the tashkeel range overlaps the digit range, so order of operations matters.
• Bidi control characters — LRM/RLM and zero-width joiners are invisible, and strip "clean" text unless removed deliberately.
• Offsets — every text tool returns start/end into the original string, so a result can be applied to the caller's own text rather than a rewritten copy.

The paid routes declare this in the 402 itself: each challenge carries a Bazaar input schema and an output example, so a client can see exactly what it is buying before it signs anything.

Endpoints

POST /text/normalize accepts a mode:

• light — lossless: whitespace, digit unification, bidi cleanup only.
• search — light + tashkeel removal + letter folding (أإآ→ا, ى→ي, ة→ه).
• index — search + punctuation stripped (use as a dedup / index key).

Try it

# 1. Any paid route answers 402 with a payment challenge
curl -i https://arabic-pipeline.arabic-pipeline.workers.dev/text/normalize \
  -H 'content-type: application/json' \
  -d '{"text":"مُحَمَّدٌ عَلِيّ","mode":"search"}'

# 2. Free routes (no payment needed)
curl https://arabic-pipeline.arabic-pipeline.workers.dev/openapi.json
curl https://arabic-pipeline.arabic-pipeline.workers.dev/

How x402 payment works (for agents)

1. Call the route with no payment header → 402 Payment Required.
2. The response carries a base64 PAYMENT-REQUIRED header and, in the body, accepts[] entries describing how to pay (network, asset, price, payTo).
3. Pick one entry, sign a payment for it, and retry the request with the PAYMENT-SIGNATURE header (base64 payment payload).
4. The server verifies and settles the payment (via the facilitator) and returns the real result plus a PAYMENT-RESPONSE header with the receipt.

The full machine-readable contract — including per-operation x-payment-info — is at GET /openapi.json.

Settlement

• Network: Base mainnet (eip155:8453), asset USDC (0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913).
• Payments are verified and settled by a public x402 facilitator (https://x402.dexter.cash — no account, no fees). The facilitator only executes the buyer's signed payment; it never holds funds.
• Settlement destination: 0x05523af5bd166816D1418Ebf7D6371e1fFB1D2e6.

PayAPI Market independently verified this deployment by buying a call and publishing the receipt: POST /text/normalize for $0.005 USDC, settled in transaction 0xfc3e42f9…73ce9c at block 53,330,178 on Base. The listing is payapi.market/api/arabic-pipeline-api.

Model endpoints

/audio/transcribe uses gemini-3.5-transcribe; /doc/extract and /doc/extract/batch use gemini-3.6-flash with a JSON response schema. Both model names are configurable (GEMINI_TRANSCRIBE_MODEL, GEMINI_VISION_MODEL). The six text endpoints are pure computation and cost nothing to serve.

A batch runs one model call per document, concurrently. At $0.10 for five documents that is $0.02 each, a third below the $0.03 single-document price; for one or two documents the single route is cheaper, and the batch route says so. A failed document is reported in place (ok: false plus error) instead of failing the whole call, so a bad scan costs one slot rather than the payment.

Local development

npm install
cp .dev.vars.example .dev.vars   # then add GEMINI_API_KEY
npm run dev                       # wrangler dev
npm test                          # vitest
npm run typecheck

Deployment

See DEPLOY.md. The short version: npm run login → npm run secret → npm run deploy → register the listing.

Two Workers are deployed from this repo:

npm run deploy              # REST API  -> arabic-pipeline.arabic-pipeline.workers.dev
npx wrangler deploy -c wrangler.mcp.toml   # MCP server -> arabic-pipeline-mcp.arabic-pipeline.workers.dev

Licence

Proprietary — all rights reserved. The source is public so the service can be audited; it is not licensed for reuse or redistribution. The API itself is sold per call, not licensed.