Trade with AI
Trade on-chain from your AI assistant
Connect Claude, ChatGPT, Cursor or any MCP client to Epsilon and place limit, stop-loss, DCA and trailing-stop orders on Robinhood Chain in plain language — non-custodial, under caps you set, and executed by keepers long after the chat is closed.
Why trade through an assistant at all
Because the hard part of trading on-chain was never the swap — it was everything around it: finding the right contract, sizing in raw units, reading a route, remembering to set the stop, checking back. An assistant with Epsilon's tools does the bookkeeping, shows you the exact numbers, and asks before anything is signed. You keep the decisions; it keeps the discipline.
What makes Epsilon's MCP different
Orders that outlive the chat
Limit, stop, DCA and trailing orders are on-chain, keeper-executed. Most agent tooling can only schedule a trade while a process is running; ours fills at 3 a.m. with your laptop shut.
Non-custodial, twice over
Orders are EIP-712 signatures from your wallet. Locally the key never leaves your machine; hosted, you approve per order or grant a capped signer. Epsilon never holds funds.
Caps the model can't argue with
Per-order and daily limits, allowed order types and tokens — enforced below the model, failing closed. A persuasive prompt cannot raise them.
Quotes that are real
Routes are built and simulated before they are served; your assistant sees the verified output, not a venue's promise.
Memecoins and tokenized stocks
The same five order types on CASHCAT and on TSLA, SPY or treasuries — with eligibility surfaced before any Stock Token purchase.
Your agent earns your rank
Trades signed by your wallet earn your points, streak and fee discount whether you or your agent placed them. Builders routing others' flow earn the rev-share leg.
What people do with it
Protect a position
“Put a stop loss 12% under the market on my CASHCAT and trail it.” The stop is signed by your wallet and watched by keepers — not by a chat tab you have to keep open.
DCA without a spreadsheet
“DCA $300 into TSLA over two weeks, every 12 hours.” One signature, 28 fills at the best route each time, cancel any time with one sentence.
Research, then act
Ask for the top tokens by volume, pull a token’s card and order book, quote the route with real fees — then place the limit order in the same conversation.
Portfolio check-ins
“What’s open on my wallet and how far is each order from filling?” Read-only, keyless, safe to run on a schedule.
Ladders and brackets
Scale into a position with a ladder of limits, or bracket a buy with a take-profit limit and a stop — the assistant builds each order, you confirm each one.
Build an agent
Same tools, from code: the TypeScript and Python SDKs sign orders the way the MCP does, webhooks push fills, and routed flow earns you the on-chain rev-share leg.
How it works
- 1. Connect. Add Epsilon to your client — one click for Cursor and VS Code, a one-liner for Claude Code, Codex and Gemini, a URL for ChatGPT and Grok.
- 2. Ask. Tokens are resolved by name, amounts in dollars, prices in USDG. The assistant quotes, preflights executability, and lays out amount, route, fee and minimum output.
- 3. Confirm. Nothing is signed on the first pass. You approve; the order is signed by your wallet and submitted. Resting orders are then the keepers' job until they fill, expire or you cancel.
Connect your client
| Client | How |
|---|---|
| Cursor | One click from cursor.directory (rules + skills + MCP), or the install deeplinks in the guide. |
| Claude Code | claude mcp add epsilon -- npx -y @epsilon-exchange/mcp |
| Claude Desktop | Settings → Developer → add the local package to claude_desktop_config.json. |
| ChatGPT | Developer Mode → add the hosted URL https://api.epsilon.exchange/mcp (read-only). |
| VS Code (Copilot) | One-click vscode:mcp/install link or .vscode/mcp.json. |
| Codex | codex mcp add epsilon -- npx -y @epsilon-exchange/mcp |
| Gemini CLI | gemini mcp add epsilon npx -y @epsilon-exchange/mcp |
| Grok · anything else | Custom connector → https://api.epsilon.exchange/mcp, or any stdio client → npx -y @epsilon-exchange/mcp. |
Two servers, one tool set: the hosted server (https://api.epsilon.exchange/mcp) is keyless and read-only — quotes, routes, order books, order status, market stats, docs. The local package (@epsilon-exchange/mcp) adds order placement and cancellation, signing with a wallet you configure. Full per-client setup, env vars and the safety model are in the developer guide.
Safety model
- Read-only by default. The hosted server cannot move funds. Every tool is annotated read-only/non-destructive so clients can show the right warnings.
- Spend cap, fail-closed. The local package refuses any order above
EPSILON_MAX_ORDER_USD; with no price available the order is rejected, not waved through. - Explicit confirmation.
EPSILON_REQUIRE_CONFIRMmakes every placement a two-step: preview with exact numbers, then confirm. - Use a dedicated trading wallet with the local package — fund it with what the agent may trade, nothing more.
- Coming to the hosted server: sign-in with your Epsilon account (OAuth), per-order approval links signed in the app, and capped agent grants — the same delegation the Epsilon Telegram bot uses.
For developers
The MCP is a thin layer over the public API. Keys and usage at developers.epsilon.exchange; TypeScript and Python SDKs sign orders byte-for-byte like the indexer expects; webhooks push fills; a referrer on your orders pays you the rev-share leg on-chain, permissionlessly. Agents that read documentation themselves get llms.txt, an OpenAPI spec and an AGENTS.md block.
FAQ
Does Epsilon hold my funds when I trade from an AI assistant?
No. Every Epsilon order is an EIP-712 message signed by your wallet, not a deposit. Tokens stay in your wallet until the order fills. With the local MCP package the signing key never leaves your machine; with the hosted server, writes are either approved by you per order or signed within caps you granted in the Epsilon app. Epsilon never has custody.
What can I actually ask my assistant to do?
Search tokens, quote and route a swap, check a position, and place or cancel market, limit, stop-loss, DCA and trailing-stop orders. Resting orders are executed by on-chain keepers, so a stop you set from a chat on Monday still protects you on Friday with the chat closed.
Which AI clients work?
Any Model Context Protocol client: Claude Desktop and Claude Code, ChatGPT (Developer Mode), Cursor, VS Code Copilot, Codex, Gemini CLI, Grok, and any agent framework that speaks MCP or plain HTTP. One-click installs for Cursor and VS Code are on the developers site.
Is the hosted MCP free?
Market data, quotes, routes, order status and docs search are free and need no API key. Higher rate limits come with a free developer key. Order placement from the hosted server is rolling out behind sign-in and user-set caps; the local package can place orders today.
How do spending caps work?
In the local package you set EPSILON_MAX_ORDER_USD and the server refuses anything above it, failing closed. Optional explicit confirmation adds a human step before any order is submitted. On the hosted server, caps are set once in the Epsilon app (per order, per day, allowed order types and tokens) and enforced server-side on every call, independent of what the model asks for.
Can my agent trade tokenized stocks?
Yes — Robinhood Chain tokenized stocks and treasuries trade on Epsilon alongside memecoins and majors, with the same order types. Buying Stock Tokens is subject to the issuer’s eligibility restrictions, which the tools surface before any purchase.
Do agent trades earn Epsilon points and rank?
Trades signed by your wallet earn your account’s points, streaks and rank discounts regardless of whether a human or your agent placed them — points always follow the maker. Developers routing other people’s flow earn the on-chain rev-share leg instead.
What about more chains?
Epsilon runs on Robinhood Chain today. Every tool is chain-aware by design; additional chains are a configuration step on the roadmap, not a rewrite.
Start
Add Epsilon to Cursor · connect another client · product docs · open the app.