MULTI/DEX — the world's most advanced DeFi

Developers Showcase
22

I used ChatGPT Pro + the ICP MCP connector against MULTI/DEX during the final stretch of Season I. A few early lessons from the experiment:

First, the basic integration works. After enabling AI access in Internet Identity, adding https://mcp.internetcomputer.org/mcp as a custom ChatGPT plugin with OAuth, and starting a new chat, ChatGPT was able to discover ICP MCP tools, find my MULTI/DEX account, and read balances/open positions.

The setup path was not totally obvious at first. In ChatGPT, the relevant flow was Settings → Plugins → Developer mode → add plugin → OAuth → MCP URL. Also, the plugin only appeared usable in new chats after setup.

The biggest surprise was autonomy. Because the competition leaderboard may involve real prizes, ChatGPT would not do discretionary “go trade for me” behavior, even though MULTI/DEX itself is currently simulation/play-mode. It would, however, accept a deterministic rule set with explicit guardrails: markets, leverage, max equity per position, stop conditions, and no actions outside MULTI/DEX.

We started this experiment very late in Season I, with only a couple days left before the July 31 end date, so this was more of a live integration test than a serious leaderboard attempt. One practical lesson is that agent-style trading needs a durable execution loop. A browser/chat-driven heartbeat is useful for experimentation, but a more serious setup would probably need server-side scheduling, a canister-side agent, or some other persistent execution model.

A market-structure lesson: the strategy repeatedly found valid short signals, but trade previews were blocked by borrow caps across BTC-ICPUSD, ICP-ICPUSD, ETH-ICPUSD, and SOL-ICPUSD. That seems like the platform enforcing risk controls correctly, but it also means agents need easy access to borrow availability/cap data before deciding what markets to evaluate.

Suggestions from the experiment:

  1. Add an “agent setup” guide for ChatGPT/Claude MCP users.
  2. Make competition status, scoring, end time, eligibility, and prize rules directly queryable.
  3. Expose borrow caps/availability as first-class read-only data.
  4. Return structured preview failure reasons that are easy for agents to act on.
  5. Consider a canister-native or hosted heartbeat option for agent strategies.
  6. Provide example safe agent policies: read-only mode, proposal-only mode, deterministic rule-set mode, and full actions mode.

Overall: the experience was rough around the edges, but genuinely interesting. MULTI/DEX plus ICP MCP is already enough to show what “AI-controllable DeFi” might feel like. I’m looking forward to Season II, especially if it gives the community more time to test strategies, agent workflows, and the developer surface around MULTI/DEX.

23

I would just like to point out that Toko is more advanced DeFi than multi/dex. Just hate it when people make fake claims about stuff like this.

24

The fair description is that MULTI/DEX may be one of the most elaborate on-chain exchange demos yet produced on ICP. “The world’s most advanced DeFi” remains unproven until it has real assets, real liquidity, publicly auditable source, independently reviewed security, verifiable deployments, ownerless governance and a meaningful adversarial production history.

They have built a very flashy car. We have built the engine platform, factory, fleet-management system, diagnostics, test track and a substantial vehicle on top of it.

Respect the product—but retire the slogan.

25

Until we see Toko, take it easy bud. We understand that you are excited about your new project, but we have yet to see Toko in action.

26

You will all kneel to the power that is Toko. All of you.

27

I have a nasty cramp in my leg.

Kneeling will be hard for me.

28

Thank you everyone for trying this out and for playing. Congrats to the winners.
A special thanks for all the feedback, much appreciated. Season/ Phase 1 ended, you can check out the **Provisional** leaderboard here.

While Phase II Competition is in progress, make sure you take a look at the CODE. Please report vulnerabilities privately.

Good luck and don’t forget to share.

29

Lets all kneel to the power if only 1 of your plans gets ready and released.

30

Do you know where to submit privately the issues, we spotted a few things that need private reporting, we were honestly planning to relax this weekend but it’s rare we get to see new Motoko code…the advisory page shows 404

31

Please use multidex@dfinity.org until Github reporting will open.

32

They will get a lot of mails than, DEX just keep deleting funds. It does not return margin cancelled unfilled orders and same with filled orders. AI keep saying those funds on my main balance, but they are not, than it mixes up margin order with spot order, it does not understand if i want to open short at 2.19, instead it opens short at current price and position is only 2.19 ICP.

I do wonder how i can close short position when it just deleted my free usd balance, hell no, even if i somehow can close it, it will delete entire usd amount.

RIP MultiDEX.

33

That’s really impractical, if you need to reproduce and attach test files etc. surely they should have a way to do on GitHub

34

I do wonder what is the goal of trading rewards. People can make 10 gmail accounts, open max long, than 9 accounts buy on spot with 900k usd, close long or open short, dump all on 9 accounts and repeat. When test period ends, there wont be much people trading as platform is confusing and buggy.

35

Anyway, there’s a bit of confusion. The SECURITY.md file in the repository says the opposite of the tweet: “There is currently no bug bounty” and asks for private reporting via GitHub Security Advisory on dfinity/multidex, not by email. The Dominic’s tweet on the other hand, talks about a bug bounty and points people to multidex@dfinity.org. I think dfinity should clarify this a bit better.

36

Even though I don’t personally agree with this type of trading contest, I respect their decision to run initiatives like this. That said, the code should be open source first, together with a reproducible build, before launching the competition. That would allow the community to verify that everything is fair and transparent. They should also exclude from eligibility any accounts that were already active before the contest was announced, if there were any, as some people have suggested. That would go a long way toward addressing concerns about fairness and transparency.

37

Using trading automation bot, does that count as cheating?

38

You’ve been hanging there a while, mate.

But you were not paying attention.

39

Yea, its like playing game where everyone has cheats and only me playing without it. :face_with_symbols_on_mouth:

40

Five of our open source libraries are already ready for you to use. Public-facing applications just require a lot more work and polish.

41

Yeah, I think you’re right. Unless there are strong safeguards against that kind of behavior, it seems like the rewards could be gamed pretty easily.

And if the platform is still confusing and buggy, it’s hard to see how those incentives alone will lead to long-term user activity.