Give your AI agent the ability to transact on the blockchain -- without giving it the keys. kova is a TypeScript SDK that enforces policy constraints on every transaction, so autonomous agents can spend within the rules you define.
The AgentWallet class is the single entry point your AI agent interacts with. It orchestrates signing, policy enforcement, chain interaction, and audit logging behind one simple execute() call. Your agent sees tool schemas; your server holds the keys.
Agents describe what they want (transfer 1 SOL to Alice), not how to do it. Five intent types -- transfer, swap, mint, stake, custom -- cover all common blockchain operations. The SDK handles instruction building, signing, and broadcasting.
Every transaction passes through an ordered chain of policy rules before execution. Spending limits, rate limits, address allowlists, time windows, and human approval gates. Fail-closed design: if anything goes wrong, the transaction is denied.
Agents interact with the wallet exclusively through the Model Context Protocol (MCP). One call to createMcpServer() exposes all wallet tools to any MCP-compatible AI agent -- Claude, OpenAI, LangChain, or custom. Policy enforcement is transparent.
High-value transactions can require human sign-off before execution. Pluggable approval channels (callback-based or webhook-based) send approval requests with configurable thresholds, timeouts, and HMAC-signed payloads. No response means denial.
Every policy decision and transaction is recorded in a SHA-256 hash chain with HMAC integrity verification. Each entry references the previous hash, making it impossible to alter history undetected. A circuit breaker halts all transactions if audit logging fails.
Full Solana support: native SOL transfers and SPL token transfers. Built-in Pyth oracle integration for on-chain price feeds with configurable staleness and confidence thresholds. Chain-agnostic architecture means more chains can be added without changing your code.
Written entirely in TypeScript with strict mode enabled. Full type safety across intents, policies, results, and tool definitions. Every function parameter, return type, and configuration object is typed so your editor catches mistakes before your code runs.