# Deep First Search: safe x402 payments for AI agents

**Safe rails for the agent economy.** *Search deep. Reveal only what you choose.*

AI agents now pay for things over x402. Every one of those payments is public, and any agent can be talked into paying the wrong party. Deep First Search makes the model **propose** payments while the owner's signed policy **decides**.

[![npm](https://img.shields.io/npm/v/@deepfirstsearch/agent-pay?label=agent-pay)](https://www.npmjs.com/package/@deepfirstsearch/agent-pay) [![MCP](https://img.shields.io/npm/v/@deepfirstsearch/agent-pay-mcp?label=agent-pay-mcp)](https://www.npmjs.com/package/@deepfirstsearch/agent-pay-mcp) [![CI](https://github.com/DeepFirstHQ/deepfirstsearch/actions/workflows/ci.yml/badge.svg)](https://github.com/DeepFirstHQ/deepfirstsearch/actions/workflows/ci.yml) [![License: MIT](https://img.shields.io/badge/license-MIT-blue)](LICENSE)

## Try it in 10 seconds

No keys, no wallet, no chain: a local x402 merchant plays honest and hostile, and you watch the agent pay for data and refuse four attacks.

```bash
npx @deepfirstsearch/agent-pay demo
```

## Pick your path

| You use | Install | Guide |
|---|---|---|
| Claude Desktop, Claude Code, Cursor, OpenClaw, any MCP client | `npx @deepfirstsearch/agent-pay-mcp ./config.json` | [integrations/mcp](integrations/mcp/README.md) |
| Vercel AI SDK | `npm i @deepfirstsearch/agent-pay-ai-sdk @deepfirstsearch/agent-pay ai zod viem` | [integrations/ai-sdk](integrations/ai-sdk/README.md) |
| LangChain.js / LangGraph | `npm i @deepfirstsearch/agent-pay-langchain @deepfirstsearch/agent-pay @langchain/core @langchain/langgraph viem` | [integrations/langchain](integrations/langchain/README.md) |
| Coinbase AgentKit | `npm i @deepfirstsearch/agent-pay-agentkit @deepfirstsearch/agent-pay @coinbase/agentkit viem@2.38.3 zod@3` | [integrations/agentkit](integrations/agentkit/README.md) |
| Your own agent, any wallet | `npm i @deepfirstsearch/agent-pay viem` | [sdk](sdk/README.md) |
| A Turnkey-held key | `npm i @deepfirstsearch/agent-pay-turnkey @deepfirstsearch/agent-pay @turnkey/sdk-server viem` | [Turnkey](integrations/turnkey/README.md) |
| A Privy server wallet | `npm i @deepfirstsearch/agent-pay-privy @deepfirstsearch/agent-pay @privy-io/node viem` | [Privy](integrations/privy/README.md) |
| A Coinbase CDP Server Wallet | `npm i @deepfirstsearch/agent-pay-cdp @deepfirstsearch/agent-pay @coinbase/cdp-sdk viem` | [CDP](integrations/cdp/README.md) |
| Agents that place real orders (food, shopping) | `npm i @deepfirstsearch/order-guard` | [integrations/order-guard](integrations/order-guard/README.md) |
| On-chain budgets on Base | `npx @deepfirstsearch/agent-pay owner help` | [owner CLI](sdk/README.md#owner-cli-vault-and-budgets-in-three-commands) |

The integrations take the SDK as a peer dependency, so your app always uses a single copy of it. Step-by-step guides for 34 x402 merchants, each tested with a real payment on Base mainnet, are in [docs/integrations](docs/integrations/).

Full developer guide: **[deepfirstsearch.com/developers](https://deepfirstsearch.com/developers.html)** · building a wallet, framework or platform? **[Partners](https://deepfirstsearch.com/partners.html)**.

## What's in this repository

| Component | What it does | Status |
|---|---|---|
| **MCP server** (`integrations/mcp/`) | Model Context Protocol server: Claude Desktop, Claude Code, Cursor and any MCP client get `paid_fetch`, `list_merchants` and `budget_status` tools that pay x402 APIs inside the owner's budget; x402 v1 merchants opt in per merchant (`x402Versions`), and an unreadable receipt is confirmed on-chain instead of resent (`confirmOnChain`, on by default) | ✅ `npx @deepfirstsearch/agent-pay-mcp ./config.json` (0.2.1) · 24 tests · tested with a real Claude session on Base Sepolia |
| **Agent Safe** (`contracts/`) | On-chain vault: owner-signed, timelocked, per-merchant budgets; the agent key can spend but never widen | ✅ 98 tests (fuzzing, invariants, properties, regressions, Base mainnet fork with real USDC), two internal reviews · live on Base mainnet (beta) |
| **agent-pay SDK** (`sdk/`) | x402 v2 client (v1 opt-in per merchant) with a policy engine outside the model, prompt-injection guards, stable refusal and block codes, never a second signature while the first might have paid, per-merchant payers, owner CLI | ✅ 293 tests (283 run offline; 10 need a Base mainnet fork or RPC) incl. on-chain loops and the official x402 facilitator · `npx @deepfirstsearch/agent-pay owner …` |
| **Framework tools** (`integrations/`) | Vercel AI SDK `paidFetchTool`, LangChain/LangGraph `createPaidFetchTool`, Coinbase AgentKit `agentPayActionProvider`; wallet payers `turnkeyPayer`, `privyPayer`, `cdpPayer`; `order-guard` for agents that place real orders | ✅ on npm |
| **$DEPTH** (`contracts/`) | 1B fixed supply, no mint function, no owner; fees burned through a fee jar + firepit | ✅ Implemented · not launched |
| **Website** (`web/`) | Site, whitepaper, tokenomics, partner kit, legal pages; strict CSP, no cookies; only third party is cookieless Cloudflare Web Analytics | ✅ |

## MCP server: let Claude or Cursor pay x402 APIs safely

`@deepfirstsearch/agent-pay-mcp` is a stdio MCP server. The model gets three tools and nothing that can move money on its own terms:

| MCP tool | What it does |
|---|---|
| `paid_fetch(url, method?, body?, contentType?)` | Fetches a URL; if it answers `402 Payment Required`, pays in USDC on Base only when the merchant, price and budget match your config |
| `list_merchants()` | Merchants the agent may pay, their prices and the remaining budget |
| `budget_status()` | Remaining budget per merchant, plan window, vault balance |

```json
{ "mcpServers": { "agent-pay": {
    "command": "npx", "args": ["-y", "@deepfirstsearch/agent-pay-mcp", "/path/config.json"],
    "env": { "AGENT_PAY_AGENT_KEY": "0x…", "AGENT_PAY_BURNER_SEED": "0x…" } } } }
```

Payees, prices and limits come from the config file and the owner-signed on-chain budget, never from the model or a 402 response, so a prompt-injected agent can't overspend or pay an attacker. Setup, config reference and a 5-minute Base Sepolia demo: [integrations/mcp](integrations/mcp/README.md).

> Status: pre-launch. **Live on Base Sepolia**, with a [real x402 payment settled](https://sepolia.basescan.org/tx/0x28c60778fcc40110440ec7fb7c944ad28fa90d5a5c26f639a5001c2736d85e5b) (see [deployments](docs/DEPLOYMENTS.md)); **unaudited beta on Base mainnet** since 2026-10-06 (small amounts, our own funds first, then invited design partners) while we arrange an independent audit; the public launch comes after the audit. No token exists.

## Develop on this repository

```bash
# Contracts (needs Foundry: curl -L https://foundry.paradigm.xyz | bash && foundryup)
git submodule update --init --recursive
cd contracts && forge test

# SDK
cd ../sdk && npm ci && npm test && npm run demo

# Website
cd ../web && npm ci && npm run dev
```

## Documents

| | |
|---|---|
| [Whitepaper](docs/WHITEPAPER.md) | Formal design, privacy model, risks (EN; structured for MiCA Annex I) |
| [Tokenomics](docs/TOKENOMICS.md) | Supply, allocation, Firepit burn, vesting (EN) |
| [Decisions](docs/DECISIONS.md) | ADRs: ticker, chain, burn, vesting, stand… (ES) |
| [Deployments](docs/DEPLOYMENTS.md) | Contract addresses on Base mainnet (unaudited beta) and Base Sepolia, plus the live demo vault |
| [Audit scope](docs/audit/SCOPE.md) | Scope, nSLOC, roles, known issues and questions for reviewers |
| [Airdrop tool](tools/airdrop/build.mjs) | Builds the airdrop Merkle tree and proofs; tested against the contract |
| [Mainnet runbook](docs/MAINNET.md) | Safes, rehearsal, deploy, checks, guarded launch, monitoring |
| [Security assessment](docs/assessments/SECURITY.md) | Pentest findings, threat model, evidence (ES) |
| [Research](docs/research/) | Market pain, successes, failures, token models, launch mechanics (ES) |

## Key decisions
- **Ticker:** `$DEPTH`.
- **Chain:** Base.
- **Founder:** 12%, nothing for 12 months, then 36 months linear, on-chain; the vesting contract cannot be transferred and the beneficiary is a public Safe.
- **Burn:** burn-to-claim fee jar (no swaps, oracles or admins), fed only by Agent Safe, SDK and inference fees, never by privacy-pool fees.
- **No conference booth** until there are users; hackathons instead.
- **No own privacy pool:** integrate a third-party compliant one (ADR-011). Sales geo-blocked for the US and Argentina (ADR-012).

## Using it? Tell us

There's no telemetry in the packages, so we only know who uses them if you say so. If your agent pays with Agent Safe, even once on testnet, tell us in [this discussion](https://github.com/DeepFirstHQ/deepfirstsearch/discussions/39): what it pays for, what got refused that shouldn't have, and what's missing. We'll test your sellers live and write a verified guide for any that's missing.

## Security
See [SECURITY.md](SECURITY.md) to report vulnerabilities.

## Contact

Built by Nicolas Tursi. General contact: hello@deepfirstsearch.com. Security reports: see [SECURITY.md](SECURITY.md) or write to security@deepfirstsearch.com.

## License
MIT (contracts and SDK).