AI Agents Can Now Trade Tokens on OneDex Without an API Key or Fees
OneDex, a decentralized token exchange, has opened an MCP server to AI agents: a single connector through which an assistant can fetch prices and buy coins on its own. No API key (the access credential for the service) is required, and according to the project, there is no extra fee. Tell the agent "Buy ONE for 0.3 EGLD" and it will find a route, check the trade and sign it locally. Set limits on amounts and tokens in code, since a model may interpret a prompt in its own way.
AI-processed from X: @OneDex_X; edited by Hamidun News
Autonomous agents have long been writing code and booking flights, but they are usually not allowed anywhere near money: few people want to hand a neural network the keys to their wallet. The OneDex project proposes a scheme in which the agent trades on its own while the keys stay with the owner. OneDex is a DEX aggregator, meaning a decentralized token exchange: it searches for which pools offer a better rate for swapping one coin for another, without a bank or an intermediary exchange.
It has now opened a remote MCP server. MCP (Model Context Protocol) is a common connector through which AI assistants such as Claude or ChatGPT attach external tools. Access goes through a single address, with no API key (a special password for the service) and, according to the project, no additional fee beyond the regular network charge.
In simple terms: you write to the agent "Buy ONE for 0.3 EGLD." ONE is the token of OneDex itself, and EGLD is the main coin of the MultiversX network. The agent then works through the chain on its own: it requests a quote, assembles the transaction, checks it, simulates it, signs it locally, sends it and waits for confirmation.
What was actually launched
Through MCP, OneDex exposed six tools. No authorization is needed and no API key is required. The service handles finding a route between pools and preparing the transaction. The agent signs it on its own side, so OneDex never sees the private keys.
A separate safeguard is the minOut parameter. This is the minimum number of tokens you are willing to receive from the exchange. The check happens directly on the blockchain: if the price moves too far during the trade, the transaction will not go through.
How to build an autonomous agent
The project describes five steps, which is essentially basic hygiene for any agent that handles money:
- Create a separate wallet with a small budget, not linked to your main funds.
- Store its key in GitHub Secrets or in the server's environment variables. Not in the code and not in a chat with the AI.
- Set the rules in code: a maximum per trade, a list of allowed tokens, a slippage limit (the permissible price deviation) and a daily budget.
- Give the command in plain language.
- From there, the agent does everything on its own.
An important detail: the limits live in your code, not in the prompt. A request like "don't spend more than such-and-such an amount" may be interpreted by the model in its own way, whereas a programmatic check will not allow it.
Manual mode for Claude and other clients
There is also a cautious option without autonomy. In Claude, you add a custom connector, paste the server address and choose access without authorization. After connecting, six tools appear. If you leave the "Needs approval" mode on, each step has to be confirmed by hand.
In this mode, the assistant is given only the public wallet address, which starts with erd1…, and is asked to perform the swap and simulation. In response, it provides a link for signing. You open it, check the recipient and amount in your own wallet, and sign yourself. The 24-word seed phrase, the key or the password must never be passed to the assistant under any circumstances.
The same address is added as a remote MCP server in Cursor, Claude Code, ChatGPT and any other client that supports MCP. The project has placed exact configurations and a TypeScript example in its documentation.
What practitioners should watch
The news is modest: it is an announcement on social media, with no independent tests and no data on how agents actually trade. The claim of no fees should be verified on your own trades.
Still, it shows a working pattern for anyone building agents that take payment actions:
- separate roles: the service prepares the operation, the agent signs locally;
- keep a separate wallet with a small budget, so a mistake costs little;
- set limits in code, not in the text of the request;
- verify the result at the network level, as minOut does;
- first run the agent with manual confirmation, and only then remove those restrictions.
Risks remain that the announcement does not remove. An agent can receive a malicious instruction from external data. Limits protect against large losses but not against poor decisions made within the budget. For product teams the takeaway is simple: MCP is becoming the standard way to connect real services to agents, and security questions need to be addressed at the design stage, not after the first incident.
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026A swimming team captain handed ChatGPT three scheduled tasks and cut her routine work by a multiple
- Nov 21, 2026OpenAI GPT-6 Cyber Found About 35 Flaws in iPhone, Mac and Other Apple Devices
- Nov 21, 202622 Claude Agents Spent Six Weeks on Black Hole Problems — and Solved Two
Key Questions Answered
What was actually launched?
Through MCP, OneDex exposed six tools. No authorization is needed and no API key is required. The service handles finding a route between pools and preparing the transaction. The agent signs it on its own side, so OneDex never sees the private keys. A separate safeguard is the minOut parameter. This is the minimum number of tokens you are willing to receive from the exchange. The check happens directly on the blockchain: if the price moves too far during the trade, the transaction will not go through.
How to build an autonomous agent?
The project describes five steps, which is essentially basic hygiene for any agent that handles money: Create a separate wallet with a small budget, not linked to your main funds. Store its key in GitHub Secrets or in the server's environment variables. Not in the code and not in a chat with the AI. Set the rules in code: a maximum per trade, a list of allowed tokens, a slippage limit (the permissible price deviation) and a daily budget. Give the command in plain language.
What are the key details on Manual mode for Claude and other clients?
There is also a cautious option without autonomy. In Claude, you add a custom connector, paste the server address and choose access without authorization. After connecting, six tools appear. If you leave the "Needs approval" mode on, each step has to be confirmed by hand. In this mode, the assistant is given only the public wallet address, which starts with erd1…, and is asked to perform the swap and simulation. In response, it provides a link for signing.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.