The toolkit never holds keys.
Every transaction-building tool returns unsigned tx_bytes_base64. The host signs, the toolkit submits. A tool that holds a key is a tool that can spend money.
Suisei puts Sui behind the Model Context Protocol, so any AI agent can read, build, simulate, sign, and submit on chain. Full portfolio visibility, persistent memory, transaction safety, and boundless agent autonomy. All non-custodial.
$ claude mcp add suisei -- npx -y @suisei-mcp/mcp
Added MCP server suisei. 37 tools.
you: what is the balance of alice.sui?
sui_get_balance
{
"address": "0x9a1f...c20e",
"balance": "42.738601200",
"coin": "SUI"
}Works with any MCP client
37
tools, one install
2
packages on npm
0
keys held by the toolkit
100%
on real testnet
The difference
A human SDK hands you primitives and expects you to assemble them. Suisei hands the agent the whole Sui Stack as one-line tools, so it reasons, decides, and acts on its own.
agent-firstStructured JSON in, structured JSON out. No React, no wallet adapter, no human in the build loop.
sui_get_portfolioCoins, stakes, rewards, and total SUI exposure in a single read. The hand-joined fan-out everyone else writes does not exist here.
sui_explain_txDecode, simulate, and judge any transaction before a key ever touches it. No other Sui SDK ships this.
mnemosui_*Indexed on Sui, stored on Walrus, portable between clients. The agent remembers what it did and why.
agent_vaultOn-chain spend limits, recipient allowlists, and expiry. Hand an agent real funds without handing it your keys.
MCPClaude Desktop, Cursor, Claude web and mobile, or your own bot. Not locked to one client, not locked to Sui tooling.
Non-negotiable
Two rules hold the whole design together. They are why an agent can act on chain without ever holding the keys that move money.
Every transaction-building tool returns unsigned tx_bytes_base64. The host signs, the toolkit submits. A tool that holds a key is a tool that can spend money.
Key generation is not a tool, because that would land the secret in the model prompt and logs. Signing lives in agent-signer, a separate local process.
Portfolio snapshots. Transaction safety. Persistent agent memory. Validator comparisons. Everything agents need to act autonomously. All read-only until explicitly signed.
Resolve names, list coins, read balances, metadata, and a whole-wallet portfolio.
Inspect objects, owned sets, dynamic fields, transactions, events.
Validator sets, APYs, and a delegator stake view.
Every builder returns tx_bytes_base64. The host signs.
The toolkit is live and the canvas is open. Ship something an agent does on Sui, then put your name on it.
The toolkit ships today. The path to mainnet and a safer agentic web is already mapped.
Live now
Next