magpie-x402

Solana lending protocol — liquidity without selling your bag.

Borrow SOL against your memecoins, tokenized stocks, and RWAs — and set take-profits and stops on that same collateral. Liquidity now, upside intact: collateral that can still sell itself.

x402 payment-required API endpoints for the [Magpie Capital](https://magpie.capital) protocol. AI agents and other Solana protocols can pay-per-call for credit scores, token risk assessments, and protocol analytics — no API keys, no signups, just a Solana payment. Agents borrow programmatically: pay per call, get back an unsigned Solana transaction, sign and submit it yourself (the service holds zero keys).

Take-profit, stop-loss, ladders, trailing stops — they fire on-chain into your loan's vault, the loan stays open, and proceeds reach your wallet only when you repay. Auto-sell exits are V4 in-vault orders on supported collateral (memecoins + tokenized stocks/RWAs); they are auto-sell orders that fire on-chain through a slippage stack, not guaranteed fills.

                        AI agent
                            │
                            │  GET /api/v1/credit-score?wallet=…
                            ▼
                ┌────────────────────┐
                │ 402 Payment        │
                │ Required           │  ← scheme, recipient, amount, nonce
                └─────────┬──────────┘
                          │
                          │  Solana tx
                          ▼            (transfer to MAGPIE_PAY_TO
              ┌────────────────────┐    + memo `magpie-x402:<nonce>`)
              │ on-chain ↪ Solana  │
              └─────────┬──────────┘
                        │
                        │  retry GET with header X-Payment: <signature>
                        ▼
              ┌────────────────────┐
              │  Service verifies  │  ← amount, recipient, mint, memo nonce,
              │  payment on-chain  │    nonce not previously consumed
              └─────────┬──────────┘
                        │
                        ▼
                 { score, tier, … }

What this is (and isn't)

• ✅ An open standard implementation. x402 is HTTP 402 Payment Required, designed for AI-agent-payable APIs.
• ✅ A revenue path for protocol data. Magpie's credit oracle, token risk scores, and analytics are useful to OTHER protocols and agents — this is how they pay for that access.
• ✅ Public-data-only. Every response field corresponds to data already verifiable on-chain via solscan.io or in the magpie-bot source.
• ❌ Not custodial. This service holds no keys, signs no transactions, cannot move any user funds. Even a full RCE on this host can't drain a user — see SECURITY.md.

Endpoints

All free endpoints query the on-chain Magpie program directly and have proper Cache-Control headers so CDN edges serve repeat reads without round-tripping.

Loan and pool reads are multi-version: every loan/pool is resolved across the V1 (memecoin), V3 (RWA — on the V3 launch), and V4 (in-vault auto-sell) programs and tagged with its program_version. These reads fail soft — if one version is unreachable, the rest still return and the affected version is reported in a partial map rather than erroring the whole call. The /api/v1/agent/self-limit-close/* surface lets a borrower-agent arm, modify, cancel, and list self-owned in-vault exit orders (take-profit / stop-loss) on its own V4 loans, authenticated by an Ed25519 signed envelope where the x402 payer is also the envelope signer.

🚀 Building your first Magpie agent? Read QUICKSTART.md — zero to a working autonomous borrow agent on Solana in 10 minutes, using the typed SDK.

>

📦 TypeScript SDK (@magpieloans/magpie-agent): every action as a one-liner. Borrow, lend, liquidate, post conditional intents — no HTTP plumbing. See sdk/README.md.

>

🧩 MCP server for Claude Desktop / Cursor / Windsurf / ChatGPT desktop: one config block, 26 tools. See mcp/README.md.

>

🎯 Limit-close agent quickstart — full end-to-end walkthrough for agents that arm, monitor, and steer limit-close (TP/SL) orders on borrowers' loans. Authorization flow, every endpoint, sample code, error reference, best practices. See docs/AGENT_QUICKSTART_LIMIT_CLOSE.md.

>

👉 Just want code examples? /examples/ — 13 turn-key TypeScript agents (credit fetch, liquidation keeper, full borrow loop, conditional intent, yield agent, webhook receiver, collateral screener, equity-leverage preview, more) that talk to the live production endpoint. Each is a single file, runs with npx tsx.

>

🧩 Using Claude Desktop, Cursor, Windsurf, or ChatGPT desktop? /mcp/ ships a one-config-block MCP server that exposes the full API as native tools in your agent host. Free reads work without any keypair; paid endpoints sign x402 payments locally with a configured Solana wallet.

More paid endpoints in progress (token risk score, batch credit lookups, webhook subscriptions, MCP server, LP-side build-deposit / build-withdraw) — see MARKETING.md for the agent-distribution roadmap or open an issue if you want one prioritized.

How to call a paid endpoint

Step 1 — get the challenge

curl -i https://x402.magpie.capital/api/v1/credit-score?wallet=9UuLsJ3jf8ViBNeRcwXD53re5G3ypgfKK3s2EiMMpump

Response:

HTTP/2 402
X-Payment-Required-Scheme: x402/solana/v1
X-Payment-Required-Amount: 1000000
X-Payment-Required-Recipient: <MAGPIE_PAY_TO pubkey>
X-Payment-Required-Nonce: a1b2c3...
X-Payment-Required-Memo: magpie-x402:a1b2c3...

{
  "error": "payment_required",
  "scheme": "x402/solana/v1",
  "payTo": "...",
  "amountLamports": "1000000",
  "nonce": "a1b2c3...",
  "memo": "magpie-x402:a1b2c3...",
  "instructions": "Send 1000000 lamports of SOL to ... with memo 'magpie-x402:...', then retry with header X-PAYMENT: <tx_signature>"
}

Step 2 — pay on Solana

Send a SystemProgram::transfer (or SPL Token transfer) for the exact amount to the recipient pubkey, with the memo instruction containing the challenge string. Confirm.

Step 3 — retry with the signature

curl -i \
  -H "X-Payment: <your_tx_signature>" \
  https://x402.magpie.capital/api/v1/credit-score?wallet=9UuLsJ3jf8ViBNeRcwXD53re5G3ypgfKK3s2EiMMpump

Response:

{
  "wallet": "9UuLsJ...",
  "score": 712,
  "tier": "gold",
  "range": { "min": 300, "max": 850 },
  "benefits": { "maxLtvPercent": 35, "minFeeRate": 0.0125, "maxDurationDays": 14 },
  "source": "magpie-credit-oracle"
}

Local development

git clone [email protected]:magpiecapital/magpie-x402.git
cd magpie-x402
cp .env.example .env       # fill in MAGPIE_PAY_TO + MAGPIE_LENDER_PUBKEY
npm install
npm run dev                 # http://localhost:8402

Deploy to Vercel (1-click)

The repo is structured for Vercel serverless out of the box:

• api/index.ts — Vercel-native entry, handles every request via hono/vercel adapter
• vercel.json — routes all paths to /api
• src/app.ts — the shared Hono app (also used by local dev)
• src/index.ts — local-dev Node server (NOT used by Vercel)

Step-by-step

1. In the Vercel dashboard, click "Add New" → "Project"
2. Import `magpiecapital/magpie-x402` from GitHub
3. Framework Preset: "Other" (Vercel auto-detects via vercel.json)
4. Set Environment Variables (Production + Preview + Development):

• MAGPIE_PAY_TO — your Solana treasury pubkey
• MAGPIE_LENDER_PUBKEY — the Magpie lender authority
• SOLANA_RPC_URL — use a paid Helius/Triton/QuickNode URL for speed (public RPC will rate-limit you)
• CORS_ORIGINS — comma-separated allowlist (use * only during testing)
• RATE_LIMIT_PER_MIN and RATE_LIMIT_PER_HOUR — tune per traffic

1. Click Deploy. First build runs npm install && npm run build (or just npm install since Vercel handles TS) — ~1 minute.
2. Test: curl https://<your-deployment>.vercel.app/api/v1/pool

Custom domain

After the first deploy, in Settings → Domains attach x402.magpie.capital (or any subdomain you control). Vercel handles the TLS cert automatically.

Why Node runtime, not Edge

@solana/web3.js depends on Buffer + Node crypto primitives that don't ship in Vercel's Edge runtime. Once @solana/kit (web3.js v2) stabilizes for Edge, switching gets us another latency win — but Node serverless is plenty fast for v0 (cold start ~150–250ms, warm <10ms server-side).

Multi-instance considerations (caveat for high-scale deploys)

The in-memory nonce store (in src/middleware/x402.ts) and rate-limit buckets are per-instance. At low traffic, Vercel runs a single warm instance and this works fine. At high concurrency, Vercel scales horizontally — a payment challenge issued by instance A might fail validation on instance B because B hasn't seen the nonce.

When that becomes a problem (it isn't for v0 — but if you push >20 req/s sustained, plan for it):

• Option 1: HMAC-sign the nonces with a server secret. Stateless, infinitely scalable, no external dependency.
• Option 2: Wire Vercel KV (Upstash Redis) for shared state. Drop-in via @vercel/kv.

Both paths sketched in SECURITY.md under "production hardening".

Security

See SECURITY.md for the full posture. Highlights:

• No keys ever stored or signed. This service verifies incoming payments only.
• No secrets in source. All config via env vars; .env is gitignored; .gitignore excludes any file matching *keypair*.json, *.pem, *.key, etc.
• Replay-resistant. Single-use nonces bound to payment memos.
• On-chain re-derivation. Amount, recipient, and mint are always re-read from the on-chain transaction — never trusted from the client.
• Rate-limited from day one — per-IP minute + hour buckets on every endpoint.
• Generic error responses — verification logic isn't leaked through error messages.

To report a vulnerability: open a GitHub issue tagged security, or contact us via https://magpie.capital/security.

Related repos

• magpiecapital/magpie-bot — the Telegram wallet bot + Anchor programs
• magpiecapital/magpie-site — the web app
• Magpie Capital — protocol overview + live stats

License

MIT — see LICENSE.