Solana Agent Payments Skill

The missing money layer for autonomous agents. solana-agent-kit and friends give an agent the ability to act. This skill gives it the ability to pay, get paid, and spend safely — without ever handing it an uncapped private key.

The problem this solves

AI agents increasingly need to move money: pay per-call for data/compute, settle with other agents, and charge for their own services. The pattern almost everyone ships today is a footgun:

// The god-mode default — seen in most agent kits, including solana-agent-kit quickstarts
const keypair = Keypair.fromSecretKey(bs58.decode(process.env.SOLANA_PRIVATE_KEY!));
const wallet  = new KeypairWallet(keypair); // unlimited authority over every lamport

One prompt injection, one hallucinated tool call, one loop bug — and the agent can drain the wallet to any address. There is no budget, no per-transaction cap, no destination allowlist, no audit trail, and no kill switch. This skill replaces that default with enforced, non-custodial spending controls, and wires up the two sides of agent commerce (paying and getting paid) the way the 2026 stack actually does it.

What this skill is for

Use this skill when the user asks for any of:

1. Pay — agent consumes paid resources

• "Let my agent pay for this API without an API key" / "pay per request in USDC"
• Consume x402-paywalled endpoints, MCP servers, or another agent's service
• Budget-tracked autonomous spend → pay-x402-client.md

2. Get paid — agent/API monetizes itself

• "Charge per call for my API / MCP tool / agent service"
• Add an HTTP 402 paywall (Express / Next.js / Hono / FastAPI), price routes, settle in USDC → monetize-x402-server.md

3. Spend safely — custody + enforced limits

• "Which wallet should my agent use?" → agent-wallets.md
• "Cap the agent at $X/day, allowlist destinations, add a kill switch" → spending-controls.md
• "Replace this raw SOLANA_PRIVATE_KEY with something safe" → spending-controls.md

4. Authorize — prove the agent may spend a user's money

• Human-in-the-loop approval, verifiable intent, audit trail → mandates-ap2.md

5. Settle — the USDC + Solana mechanics underneath

• Mints, decimals, ATAs, gasless (fee abstraction), finality, idempotent receipts → usdc-settlement.md

How the pieces fit (mental model)

Agent payments in 2026 are three layers, often conflated. Keep them separate:

x402 is the stablecoin settlement extension of AP2 — the A2A x402 extension carries an AP2 mandate into on-chain settlement, so crypto payments get the same audit trail card payments get. You can adopt x402 alone (most common today), or pair it with AP2 mandates when an agent spends a user's money rather than its own.

Default stack decisions (opinionated, 2026)

1. Protocol: x402 v2. Use the scoped @x402/* packages (@x402/core, @x402/svm, @x402/fetch, @x402/express|next|hono|fastify). v2 uses the PAYMENT-REQUIRED / PAYMENT-SIGNATURE / PAYMENT-RESPONSE headers and CAIP-2 networks. Treat the older single X-Payment + x402Version: 1 shape as legacy (see resources.md).
2. Currency: USDC. Solana mainnet mint EPjFWdd5AufqSSqeM2qN1xzybapC8G4wEGGkZwyTDt1v (6 decimals). x402 settles in USDC, not SOL.
3. Settlement chain: Solana mainnet. Sub-second finality, ~$0.00025 fees — ideal for micropayments. Develop on devnet first.
4. Scheme: `exact` (ExactSvmScheme) for Solana. upto and batch-settlement are EVM-only as of 2026 — do not promise them on Solana.
5. Custody: never a bare keypair in prod. Default to on-chain enforcement (Squads v4 spending limits or Crossmint smart wallets) when the threat model demands that a buggy agent cannot exceed limits. Use off-chain policy (Privy / Turnkey / Coinbase Agentic Wallets) when you want managed key infra and trust the signer service. See the decision matrix in agent-wallets.md.
6. Facilitator: don't touch the chain from your server. Use a facilitator to verify+settle: https://x402.org/facilitator (testnet), Coinbase CDP or PayAI (prod), or self-host Kora for a Solana-native, fee-abstracted facilitator. See monetize-x402-server.md.
7. Every autonomous spender gets a budget + a kill switch. Non-negotiable. See spending-controls.md.

Operating procedure

1. Classify the request

2. Pick the right agent

3. Apply the safety baseline (always)

Before any agent goes near mainnet funds, confirm all of these — the payment-safety-auditor checks them, and rules/payments.md enforces the code-level ones:

• [ ] No raw private key with unlimited authority. Funds sit behind an enforced limit (on-chain spending limit or policy-gated signer).
• [ ] Per-transaction cap and rolling budget (daily/session) are set and enforced — not just logged.
• [ ] Destination allowlist for any non-x402 transfer. (x402 pays the facilitator-validated payTo; arbitrary transfers are the dangerous path.)
• [ ] Kill switch: a single config flip / multisig action halts all spend.
• [ ] Every settlement persists its PAYMENT-RESPONSE receipt for audit.
• [ ] Secrets via env/secret manager, never committed. Wallet files gitignored.
• [ ] Devnet-tested before mainnet. Amounts handled in atomic units (USDC ×10^6).

A runnable, offline model of this baseline — the pay→settle loop with the per-tx cap, session budget, allowlist, and kill switch enforced — is in examples/devnet-x402/ (node --test, 12 passing tests).

4. Commands

Relationship to the rest of the kit

This skill is the payments + controls layer; it routes to (does not duplicate) the protocol skills:

• `solana-agent-kit` gives the agent 60+ actions. This skill governs how it pays for and is paid for those actions, and replaces its raw-keypair wallet with a capped one.
• `birdeye` x402 example is a single consumer of x402; pay-x402-client.md generalizes that pattern to any x402 resource and adds budget tracking.
• `squads` covers multisig broadly; spending-controls.md applies its spending-limit primitive specifically to autonomous agents.
• For raw token transfers, ATAs, and Token-2022 → defer to the kit's core solana-dev skill and token-2022.md.