The central problem in agentic finance is not speed or cost. It is custody. When an AI agent holds money, who holds the keys? Most implementations today rely on custodial APIs or read-only blockchain explorers, meaning the agent never truly owns its capital. The keys sit with a third party, reintroducing the counterparty risk that crypto was supposed to eliminate.
Tychi AI ships a direct answer to that question. The platform, described as an “economic layer for autonomous capital,” is a self-custody agentic wallet designed for AI agents and their human operators. It runs on Arbitrum One. Its core architecture splits the problem: private keys stay encrypted on the user’s local machine, while a hosted “brain” handles intent parsing, transaction routing, and policy evaluation. The signing never leaves the device.
That split is the differentiator. Most existing agent wallets are custodial APIs where a third party holds the keys, or standard wallets like MetaMask and Ledger that are built for human GUI interaction, not agent-native workflows. Tychi’s design is agent-first. It ships two interfaces: a Model Context Protocol (MCP) server called @tychilabs/tyi-mcp that integrates with AI hosts like Cursor, Claude Desktop, OpenClaw, and Codex, and a terminal-based REPL called @tychilabs/tyi for humans using natural language. Both operate on the same encrypted keystore at ~/.tyi.
The practical implication is that an AI agent can earn revenue from a service, hold it in its own wallet, and sign transactions without exposing its private key to any external server, including Tychi’s own hosted brain. The hosted brain processes the intent — “send 10 USDC to this address” — and evaluates policy caps at the edge before the cryptographic signing happens locally. This creates a trust layer where agent autonomy is possible without sacrificing self-custody.
The timing matters. AI agents are moving from chat interfaces to autonomous economic actors. Projects like Virtuals Protocol, Payman, and Skyfire are building agent-to-agent payment rails. But the custody question remains unresolved at scale. Tychi’s approach is not just a technical choice; it is a structural bet on how agent finance should work. If agents are to participate in DeFi, manage portfolios, or execute microtransactions autonomously, they need wallets that treat them as first-class principals, not as scripted extensions of a human’s MetaMask.
Tychi’s policy engine adds a governance layer that is missing from most standard wallets. Before any on-chain action, the system evaluates caps and guardrails defined by the human operator. This allows a human to set up a wallet via the CLI, then grant an AI agent managed access through the MCP configuration. The same ~/.tyi keystore is shared. The result is a unified human-agent wallet where the human defines the rules and the agent executes within them.
The gasless routing via the Universal Gas Framework (UGF) is a practical necessity for agent workflows. Microtransactions on Ethereum mainnet are prohibitively expensive. Tychi’s integration with Arbitrum and UGF enables gasless sends for native ETH and ERC-20 tokens, removing the cost barrier for frequent agent-to-agent payments.
The product is in public beta on npm. Both @tychilabs/tyi@beta and @tychilabs/tyi-mcp@beta are available now. The documentation is clear about the limitations: currently deployed only on Arbitrum One, with expanded DeFi functionality on the roadmap. The FAQ explicitly addresses the key safety question — private keys never leave the local machine — and explains the MCP integration as the mechanism that allows any compatible AI agent to use the wallet as a native tool.
The open question is adoption. Tychi’s architecture is sound, but it enters a crowded space where incumbent wallets and custodial APIs have network effects. Developers building autonomous agents must choose whether to trust a new key management system with their agent’s capital. The split custody model reduces that risk, but it introduces a new dependency on Tychi’s hosted brain for intent parsing and policy evaluation. If the brain goes down, the agent cannot route transactions.
For AI builders, the takeaway is clear. The era of agents holding their own keys has arrived on Arbitrum, with a design that prioritizes local signing and policy-at-the-edge control. The next six months will show whether developers trust it enough to let their agents earn, hold, and spend.