say ur intent

Demo

About

Keep your AI. Keep your context. Keep your wallet. Say Ur Intent brings Sui DeFi tools and interactive transaction review into AI desktop applications. The integration targets MCP Apps-capable versions of the Codex desktop app and Claude Desktop. Users run Say Ur Intent on their own computer and connect it to their AI app through the Model Context Protocol (MCP), which lets AI applications call external tools. MCP Apps lets those tools display interactive cards inside the conversation. Users continue with their existing AI service, model and conversation without creating a separate Say Ur Intent AI account or subscription. A user can ask to see their Sui wallet assets, check a DeepBook quote, and prepare a supported swap with an explicit amount and slippage limit. Asset and review cards appear in the AI app, while protocol information is returned to the conversation. A follow-up such as "Use a smaller amount; keep the other conditions" carries the earlier choices into a new request and a new transaction review. Before requesting wallet approval, the user sees what they will spend, the expected receive, the minimum receive if execution succeeds, the selected account, estimated gas and simulation findings. A graph of the Sui programmable transaction block (PTB) shows the underlying transaction commands for further inspection. The user requests approval from the Review card and signs that individual transaction in their own wallet. Say Ur Intent checks the returned transaction and signature against the reviewed transaction and selected account before submitting once. Confirmed execution facts return to the same conversation. The user can then ask about the transaction's actual asset changes and request fresh wallet balances without manually pasting a wallet result into the chat. Unconfirmed outcomes remain explicitly unknown. DeepBook and FlowX are the current swap adapters. A shared review and wallet workflow gives Say Ur Intent a reusable utility structure for AI tools and additional supported protocol adapters. Structured proposals from other tools can currently be reviewed; executing or signing those external proposals remains future work. The AI and cards receive transaction facts and status, never executable transaction bytes or wallet signatures. Each transaction requires the user's own approval, and estimates remain distinct from confirmed chain results.

How it's made

Say Ur Intent is a TypeScript application running locally on Node.js. The Model Context Protocol SDK exposes tools that an AI application can call, and MCP Apps supplies interactive views rendered inside the conversation. The AI desktop app provides the language model and conversation context; Say Ur Intent provides Sui data, transaction preparation, review and wallet interaction. The backend uses the Sui SDK for mainnet reads, transaction construction and simulation access. Pinned DeepBook and FlowX SDKs supply protocol-specific quotes and swap builders through shared adapter contracts. This allows each adapter to handle its protocol's concrete APIs while using the same review and approval workflow. Zod validates structured inputs and evidence contracts. Token quantities remain integer strings or BigInt values, with units derived from verified metadata. SQLite is the shared source of truth for accounts, cards, reviews and transaction requests. An explicit user action in the card records the exact review revision, transaction and account before a wallet signature is requested. WalletConnect carries a sign-only request between the private backend and the approved wallet. The backend recalculates the digest from the returned transaction bytes, verifies the selected account's signature and checks the mainnet connection before submitting once. The review binds human-readable facts and simulation evidence to the same transaction digest. PTB diagrams use @zktx.io/ptb-model and Mermaid to show transaction structure. Market charts use Lightweight Charts with official DeepBook Indexer candle data. Transaction bytes and signatures remain outside model-facing responses and card display state. A central engineering challenge is handling asynchronous wallet approval and uncertain network responses without repeating a financial request. Review revisions and atomic admission keep requests tied to the user's selection. Request status is stored separately from independently observed chain effects. Reopening a card reads stored state; a lost submission response leads to observation of the same transaction digest, not another submission. Tests exercise SQLite and signature verification with controlled wallet and chain dependencies. This Continuity project extends an existing open-source Sui evidence and review toolkit. Live end-to-end verification of the revised embedded wallet workflow has not yet been completed. OpenAI Codex was also used as a development assistant for implementation, tests, technical review and documentation. The maintainer defined the product requirements and reviewed implementation plans.