Froggy

Demo

About

Froggy is a persistent workspace that gives AI agents the capabilities they need to actually get things done, instead of stopping at an answer. It combines a shared browser, a wallet with spending controls, an email inbox, Telegram notifications, monitoring, research tools, trading and task history in one place. You can use Froggy directly or connect an agent you already use through MCP. That agent can research, browse websites, watch wallets or saved items, prepare reports, receive email, use paid services and request financial actions while operating under limits it cannot change. The core idea is to separate delegating work from delegating authority. Agents can do useful work autonomously, but sensitive actions such as transfers, trades, dapp transactions, signatures and email sends can require the human. Froggy keeps the context, receipts and results together so the assistant you use can change without your ongoing work having to start over.

How it's made

Froggy is a TypeScript monorepo built on Bun with a React web app and a Bun server. The server owns durable agent runs, task history, approvals and recovery so work can continue even if the browser tab disconnects. External agents connect through a Streamable HTTP MCP server with OAuth 2.1, and there is also a task API, CLI and Telegram interface. Privy provides the user-controlled embedded EVM wallet and policy layer. We use it for wallet ownership, signing and controlled agent execution. Agents never receive the user’s private key and cannot increase their own spending limits. Hedera + x402 are used as a payment rail for purchasing Froggy credits with HBAR. The payment flow settles through Blocky402 and can record settlement evidence through HCS. Credits are deliberately separate from the user’s wallet funds: credits pay for agent work, while wallet funds are used for supported purchases and trades. The Graph powers onchain research. Froggy can query a standardized set of lending deployments and return the deployment and indexed block behind the result instead of only returning model-generated prose. Froggy also has a hosted shared Chrome environment. The agent can browse and operate websites while the person watches the same session, takes control when necessary, and hands it back afterward. The browser can expose Froggy’s wallet to supported dapps while signing remains outside hostile page content. We built a real email subsystem for the agent as well: inbound mail, attachments, drafts, delivery state and human-reviewed sending. This lets an agent receive verification emails, work with support replies and deliver generated reports. Telegram gives the agent an asynchronous way to notify the owner about updates and tasks that need attention. Monitoring and Watchlists let Froggy follow wallets, tokens and saved items over time, including Base and Robinhood Chain activity, rather than requiring the user to repeatedly check dashboards themselves. One architectural detail we cared about a lot is the security boundary: the browser package and wallet package are intentionally isolated, pages and model output cannot grant spending authority, every external wire message is decoded, and irreversible or uncertain financial operations are persisted before execution so retries do not accidentally duplicate payments.