🚀 BANK OF AI

Building a Decentralized Financial Sovereignty and Commercial Interconnection Layer for AI Agents

BANK OF AI is designed to empower AI Agents with financial sovereignty, identity authentication, skill extensibility, and automated payment capabilities. Through standardized protocols, we enable AI to autonomously earn income, pay for resources, and build credit on the blockchain—just like humans.

Currently supports TRON and BSC, with more blockchain networks coming in the future.

x402 is an open blockchain payment protocol based on the HTTP 402 Payment Required status code. It removes payment friction in M2M (machine-to-machine) interactions for AI agents.

• Payment-as-Response: Enables a “pay-before-response” mechanism without complex account registration or session management.
• Multi-Chain Support: Currently supports TRON (Mainnet/Shasta/Nile) and BSC (Mainnet/Testnet); More blockchain networks will be supported in the future.
• AI-Friendly: Supports signed payloads, allowing agents to complete settlements automatically.
• Use Cases: Pay-per-request APIs, paywalls, and automated settlement between agents.

Built on the Model Context Protocol (MCP) architecture, providing full on-chain execution capabilities for agents.

• Infrastructure Layer (MCP Server): The agent’s “hands and eyes.” Provides a unified interface to read blocks, manage wallets, execute transfers, and interact with any contract.
• Application Logic Layer (Skills): The agent’s “brain.” Uses structured instructions (SKILL.md) to guide AI in orchestrating underlying tools.
  • sunswap: Enables automated DEX swaps.
  • wallet-cli: Guides agents through wallet-cli for TRON accounts, transfers, signing, and on-chain operations.

For new setups, start with Skills and wallet-cli (see Quick Start). OpenClaw remains documented for existing integrations.

OpenClaw is a financial extension plugin developed by BANK OF AI for AI agents, aiming to become the “central bank” of the agent economy.

• Wallet Ownership: Enables agents to truly control and manage on-chain assets.
• Autonomous Spending: Agents can independently pay for computation, storage, and data resources.
• Seamless Integration: Built-in MCP server and automatic skill installer for plug-and-play setup.

Want to monetize your API, content, or services?

1. Integrate x402: Deploy server-side logic to detect payment requirements and return a 402 status code with payment instructions.
2. Publish Skills: Write a SKILL.md for your service so other AI agents can understand how to interact with it.

Want your agent to have autonomous payment and on-chain interaction capabilities?

1. Install Skills: Add the Skills your agent needs, such as wallet-cli or sunswap, from the skills repository. See the Skills quick start.
2. Set up wallet-cli: Install wallet-cli and configure an account on your own machine. See the wallet-cli quick start. Never send a private key, mnemonic, or password to an AI.
3. Start Transacting: Your agent can now pay x402 endpoints with wallet-cli x402 pay and run on-chain workflows through Skills.

Already using OpenClaw or the TRON MCP Server? Their docs remain available: OpenClaw and TRON MCP Server.

• Documentation Center: docs.bankofai.io
• Reference Examples: x402-examples
• Technical Support: Submit issues or join discussions on GitHub.

BANK OF AI — Empowering AI Agents with Financial Sovereignty.

Use Node.js 22.23.4 (.nvmrc, matching CI and the Docker build) and Yarn Classic 1.22.x. Install dependencies on the machine that runs the build; copying node_modules between Intel and Apple Silicon machines can break the native Chinese search tokenizer.

nvm install
nvm use
yarn install --frozen-lockfile
yarn lint
yarn typecheck
yarn test # math rendering, KaTeX and CSP security regressions
yarn build # builds English and Simplified Chinese, including search indexes
yarn serve

Use yarn start for English development or yarn start-zh for Chinese. OpenSSL legacy mode is no longer required.

If a build reports Failed to load native binding from @node-rs/jieba, check node -p "process.version + ' ' + process.arch". An Intel (x64) Node cannot use an Apple Silicon (arm64) tokenizer package, or vice versa. Run nvm install, nvm use, then yarn install --frozen-lockfile --force in this directory before rebuilding. Install and build in the same terminal session, and keep optional dependencies enabled. The build now checks this binding before compiling either locale and prints these recovery steps if it cannot load.

Keep all official @docusaurus/* packages on 3.10.2 together. The site uses React 18, MDX 3, TypeScript 5.9 and @easyops-cn/docusaurus-search-local 0.55.3. When changing the framework or search theme, check both locales, all three documentation sections, sidebar navigation, search results and legacy redirects. Markdown links remain strict build errors.

KaTeX is pinned to 0.19.1, including the transitive copy used by rehype-katex. The latter still declares KaTeX 0.16.x, which is affected by GHSA-238p-pmpm-9mq7. Keep this override until the upstream plugin supports a patched release, and verify inline/display equations and untrusted commands when updating it. CSS and fonts are bundled from the same installed KaTeX package.

The other Yarn resolutions apply security fixes beyond upstream dependency ranges: Tinypool 2.2.0 (worker options), serialize-javascript 7.1.4 (serialization), UUID 11.1.1 only under SockJS (buffer bounds), and minimatch 11.0.7 only under typescript-estree (glob amplification). Keep their scopes narrow and recheck a build with DOCUSAURUS_SSG_WORKER_THREAD_COUNT=2, lint, type checking and the development server when changing them. Yarn's range warnings for these reviewed overrides are expected.

The site explicitly loads the official docs, pages, debug and sitemap plugins and the classic theme instead of @docusaurus/preset-classic. This removes the unused SVGR dependency chain containing deepmerge (GHSA-gcqx-w48m-w2h4, no patched release as of November 11, 2026). SVG files remain supported as ordinary images; SVG-to-React component imports are not used by this site. Do not reintroduce the preset or SVGR without checking this dependency. Verify with yarn audit after updating dependencies.

The container assembles a minimal runtime from the official Alpine Nginx binary and its ELF library dependencies. BusyBox, its wget/ssl_client tooling, shells, package managers and unused dynamic Nginx modules stay outside the final image. The runtime keeps the original Nginx APK inventory and the installed libraries' package records so Grype still scans the software that ships. CI refreshes the runtime-files stage on every build; its Alpine packages are not reused from a cached security-update layer.

The container still listens on port 80 and writes logs to /app/logs. Its entrypoint directly runs /usr/sbin/nginx, preserves graceful SIGQUIT shutdown, and validates configuration with an exec-form nginx -t during the build. There is no docker exec ... sh; inspect logs externally or run docker run --rm IMAGE -t to check configuration.

yarn build also generates .site-security/nginx-headers.conf from the emitted HTML. Docker copies this policy outside the public document root and runs nginx -t. The CSP allows the exact inline script hashes needed for theme initialization and legacy redirects, while blocking other inline scripts, event handlers, external scripts, embedding and plugins. React/KaTeX inline styles remain allowed. Fonts and KaTeX assets are served locally; existing ReadMe-hosted screenshots are allowed only from files.readme.io.

Nginx applies the security headers to all responses, including localized 404 pages, disables version disclosure and directory listings, and returns relative redirects. HTML responses use Cache-Control: no-store. When serving the build outside this Docker image, apply the generated policy and equivalent response headers at the actual host. Security settings in this repository take effect on the live site only after deployment.

Both image-publishing workflows require the same dependency audit (all severities), lint, type checks and security regression tests. Actions are pinned to verified commit SHAs, checkout credentials are not persisted, and build jobs pull current base images. Each workflow builds and loads one image, scans its runtime packages with pinned Grype (including findings without fixes), and runs scripts/check-http-security.mjs against the running container before authenticating to Docker Hub and pushing that same image. Any reported vulnerability at negligible severity or above, scanner failure, or failed HTTP check blocks publishing. Production CDN/TLS settings still require verification after deployment.