AI agents that write, compile, deploy and verify smart contracts.

Describe a contract in plain language and a Web3GPT agent writes the Solidity, compiles it with the right solc version, deploys it to an EVM chain, pins the sources to IPFS and queues verification on the block explorer. You can use it in the browser, or call it from your own agent through a public HTTP skill API.

• App: w3gpt.ai
• Skill guide (for AI agents): w3gpt.ai/skill.md
• API reference: w3gpt.ai/api-docs
• Live deployments: w3gpt.ai/contracts · w3gpt.ai/api/stats

• Chat to deploy. Ask for an ERC-20, an NFT collection, a vesting contract or anything else. The agent writes it, using OpenZeppelin and other npm or GitHub imports when useful, and deploys it. Every deployment returns the contract address, transaction hash and explorer links, and is only reported after a successful contract-creation receipt.
• Two ways to pay gas. Agent deployments are signed by Web3GPT's own wallet, so no wallet is needed. "Deploy with wallet" compiles on the server and lets you sign the transaction with your own wallet via RainbowKit.
• Any Solidity version. The compiler is picked from the contract's pragma (0.4.11 and later). Downloaded compilers are checksum-verified.
• Verified and reproducible. Sources, ABI, bytecode and the standard JSON input are pinned to IPFS. A background job verifies each contract on Blockscout or Etherscan.
• Custom agents. Anyone signed in can create an agent with its own instructions and share it by link (w3gpt.ai/?a=<agentId>).
• Agent-to-agent API. POST /api/skill lets other agents and scripts hold a conversation with a Web3GPT agent, including deployments, with no account.
• Saved and shared chats. Signed-in users (GitHub) keep their chat history, including deployment results, and can publish a chat as a public link.

Polygon mainnet is intentionally available to agents only, not to wallet connectors. Retired testnets (e.g. Celo Alfajores) can no longer be deployed to, but their past deployments still appear on the dashboard.

Start a conversation. A new chatId is returned, and it is the secret for continuing that conversation:

curl https://w3gpt.ai/api/skill

Send messages with that chatId:

curl -X POST "https://w3gpt.ai/api/skill?chatId=<chatId>" \
  -H "Content-Type: application/json" \
  -d '{"message":"Deploy an ERC20 called Demo (DEMO) with 1,000,000 supply on Polygon Amoy"}'

{ "agentId": "agent_web3gpt", "chatId": "<chatId>", "response": "Deployed Demo at 0x… (tx 0x…) …" }

Add history=true to get the whole conversation back. The agent may ask you to confirm before it deploys; reply in the same chat. The full guide, written for AI agents, is skill.md. The OpenAPI schema is at /api-docs.

The public endpoints are metered so the shared deployer wallet and model budget can't be drained:

• Agent messages: 20 per IP and 30 per chat, every 10 minutes. Signed-in browser users get 60 per 10 minutes.
• Deployments paid by Web3GPT: 10 testnet deployments per caller per hour. On Polygon mainnet, 2 per caller per day, within a shared daily budget.
• Over a limit: you get HTTP 429 with a Retry-After header. "Deploy with wallet" has no gas limits, because you pay your own gas.

• Bun 1.4+
• An OpenAI API key
• A Redis store with a REST API (Upstash, e.g. through the Vercel marketplace integration)
• A funded EVM private key for agent deployments (testnet funds are enough to start)
• A Pinata account for IPFS pinning
• A GitHub OAuth app for sign-in

git clone https://github.com/markeljan/web3gpt.git
cd web3gpt
bun install
cp .env.example .env.local   # then fill it in, see below
bun dev

For local sign-in, create a GitHub OAuth app with the callback URL http://localhost:3000/api/auth/callback/github. Then set AUTH_GITHUB_ID, AUTH_GITHUB_SECRET and AUTH_SECRET (any long random string). Nothing else is needed: the production callback proxy only applies to this project's own OAuth app.

Required:

Recommended:

Optional limit overrides, with defaults in lib/rate-limit.ts: MAINNET_DEPLOY_DAILY_LIMIT, MAINNET_DEPLOY_DAILY_LIMIT_PER_CALLER, TESTNET_DEPLOY_HOURLY_LIMIT_PER_CALLER, RATE_LIMIT_CHAT_ANON, RATE_LIMIT_CHAT_USER, RATE_LIMIT_SKILL, RATE_LIMIT_SKILL_CHAT, RATE_LIMIT_SKILL_READ, RATE_LIMIT_COMPILE, RATE_LIMIT_FINALIZE.

.env.example lists every variable.

The app is built for Vercel. vercel.json schedules /api/cron (contract verification) every minute and allows API routes up to 300s, which a deployment turn can need. Set the environment variables above in the Vercel project.

Install dependencies and run the local checks:

bun install
bun run lint
bun run typecheck
bun run test

Pull requests should pass all three checks, as CI does. bun run test uses bun test --isolate to prevent test mocks from leaking between files.

bun dev              # dev server on :3000
bun run format       # auto-fix formatting and lint
bun run build        # production build

CI runs typecheck, tests and lint on every pull request.

browser (Next.js app) ──► /api/chat  ─┐
other agents ───────────► /api/skill ─┼─► agent (AI SDK + OpenAI) ─► tools: deployContract, resolveDomain, …
                                      │                                  │
                                      │        compile (solc by pragma) ─┤
                                      │        deploy (viem, server key) ┤
                                      │        pin sources (Pinata/IPFS) ┤
                                      └─► Redis: chats, agents, deployments, rate limits
Vercel cron ─► /api/cron ─► verify queued contracts on Blockscout / Etherscan

• app/: pages (/, /chat/[id], /share/[id], /contracts) and API routes
• components/: chat UI, sidebar, wallet and deploy components, ui/ primitives
• lib/: agents and chains (constants.ts), AI tools (tools.ts), persistence (data/), Solidity compile, deploy and verify (solidity/), rate limits
• test/: unit tests

AGENTS.md is the detailed guide to the codebase: architecture decisions, invariants and conventions. Read it before making changes, whether you're a person or a coding agent.

Please report vulnerabilities privately through GitHub security advisories, not public issues.

Bun audit reports 10 remaining advisories below. These are awaiting upstream compatible fixes or consumer migrations; no cross-major overrides are forced. This is dependency tracking, not a claim that every affected path is browser-only.

The existing axios override and postcss-selector-parser 6.1.5 override are retained. solc now resolves tmp 0.2.7 (fixes GHSA-7c78-jf6q-g5cm). A Bun scoped viem → ws 8.21.0 override fixes the nested WalletConnect viem 2.23.2 copies (GHSA-58qx-3vcg-4xpx and GHSA-96hv-2xvq-fx4p). Bun 1.4.2 supports this scope using lockfile version 3. ws 7.5.13 and the unrelated Solana ws 8.22.2 remain intact. These versions, ws 8.21.0, zod 4.6.7 and uuid 14.0.3 are outside the respective reported affected ranges. Remove overrides when upstream dependency ranges resolve fixed versions naturally.

MIT