Skip to content
Envelop
unisafev0.3.0
AI agents

Run Uniswap v4 liquidity from your AI agent

unisafe gives an agent three levels of access, and they differ by what they need from you rather than by how clever they are. The first needs nothing at all: a URL. The second is documentation. Only the third holds a key — and even then the contract, not the agent's restraint, decides what it may do.

Give your agent a URL — it reads, it cannot sign

A hosted MCP server over HTTP. No install, no keys, no wallet. Your agent gets position history and the fees those positions actually earned, who may operate a manager, which of that manager's pools pays best today, and whether moving a position is worth it. The server holds no key and cannot sign a transaction, so nothing it says can move money.

The endpoint
https://unisafe.envelop.is/mcp
Claude Code, Claude Desktop and other MCP clients
claude mcp add --transport http envelop-lp-insight https://unisafe.envelop.is/mcp
Codex — config.toml
[mcp_servers.envelop-lp-insight]
url = "https://unisafe.envelop.is/mcp"
ToolWhat it answers
get_position_historyEvery position of a manager: when it opened, its pool, its realised fees and its events.
list_operatorsWho may currently act on a manager. They can rebalance; they can never withdraw.
rank_poolsThe manager's own pools, scored by fee APR over the last complete day.
suggest_moveWhether moving a position to one of those pools is worth it, as numbers — never as calldata.

Every answer says which source served it and how current that source is: an index built from the chain with Substreams, the Envelop history service behind it, and a log scan last. An empty answer is an answer; a source that cannot answer says so instead of returning nothing.

Give your agent the manual — the skill

A folder of documentation connected to nothing: what a Manager NFT is, how a Stable manager differs from a Volatile one, what the fees are, and which steps only you can do in the browser — with the link for each. No keys, no configuration, and it is the whole story on a client that cannot run a local process. The hosted server above serves the same skill as MCP resources, so an agent given only the URL still gets the manual rather than four tool descriptions.

Codex, Cursor and other agents
npx skills add https://unisafe.envelop.is/skills/envelop-mcp-lp-latest.zip
On claude.ai and Claude Desktop, download the skill zip and upload it under Skills.

Give your agent an operator key — the local server

@envelop/mcp-lp runs on your machine and signs: allocate idle balance, recenter a position, move liquidity from one of your pools into another, claim fees and reinvest. You authorise its address with one transaction and revoke it with one. It can never withdraw funds, change the price oracle or authorise another operator — that boundary is in the contract, not in the agent's good behaviour — and its swaps and liquidity adds are checked against a Chainlink price and fail closed.

Reading moved out of it: history, operators and pool ranking come from the hosted server. The half that holds the key makes no outbound call except to the chain, so a remote service cannot hand it a transaction to sign — a property of the architecture rather than a promise in a README.

Claude Code — the skill and this server in one plugin
/plugin marketplace add https://unisafe.envelop.is/.claude-plugin/marketplace.json
/plugin install unisafe@envelop
Needs Node 20 or newer on the machine where your key lives. Setup, tools and the safety model →

What an agent can never do

Withdraw your funds to any address, change the price oracle, or authorise another operator. Those are the NFT holder's alone, enforced by the manager's own code. An operator moves liquidity around inside the manager and nothing else — and one transaction from you takes even that away.

New to unisafe? Stablecoin and volatile LP managers on Uniswap v4 explains the product itself.