MULTI/DEX — the world's most advanced DeFi

[ZackDS]

Sure you can, use Reporting a vulnerability or if you prefer more private than email multidex@dfinity.org instead.

[alfonso]

@alfonso

Quick Season II follow-up to my earlier ChatGPT Pro + ICP MCP test.

After manually funding the MULTI/DEX account with the $100,000 play balance, ChatGPT via ICP MCP can read the important account state: balance, leaderboard row, pools, positions, and closed orders.

Where I’m still blocked is the safe pre-action path. External ChatGPT MCP reads to methods such as owner-specific orders, quoteSwap, and previewOpenPosition are returning INVALID_ARGUMENT through the safe query path I’m using. Because my rule set requires a successful quote/preview before any trade, the agent correctly refuses to trade even when market signals are valid.

This may not mean assisted trading is disabled. Dominic mentioned that MULTI/DEX supports account/book analysis and trade orchestration through onsite chat and MCP, so maybe I’m hitting a connector limitation, an OQL/query limitation, or a difference between onsite chat and external ChatGPT MCP access.

Concrete questions:

1. Is external ChatGPT MCP expected to support safe pre-trade reads like quoteSwap, previewOpenPosition, getMyOrders, and whyAmIRefused today?

2. If yes, should those be callable as query-style reads, or does the connector need to use a signed update-style read for this OQL canister?

3. Is manual UI funding currently required before MCP-assisted trading, or should MCP be able to complete the normal play-funding flow?

4. Can Season II metadata, reward rules, end time, borrow caps, and preview failure reasons be exposed as first-class read-only data for agents?

Overall, the integration is promising. The account discovery and read-only monitoring already work well enough to show the shape of AI-controllable DeFi. The missing piece is a clearly documented safe action path so agents can preview, explain, and then act under deterministic user-approved rules.

Follow-up after testing the built-in MULTI/DEX portal assistant directly.

This adds useful context to my earlier external ChatGPT Pro + ICP MCP report.

The built-in portal assistant does appear to support the intended assisted-trading flow. I tested small spot trades only, with manual confirmation each time:

1. Sold 42 ICP back to ICPUSD after asking for price, fee, slippage, and confirmation.

2. Bought 2.0186 SOL with about $150 ICPUSD after a proposal/confirm card.

3. Later sold the SOL after it reached the take-profit area.

The confirmation flow worked: the assistant proposed the action, showed the method call, waited for confirmation, then submitted the order. That is a meaningful distinction from the external ChatGPT MCP path, where quoteSwap, previewOpenPosition, getMyOrders, and similar safe pre-action reads still return INVALID_ARGUMENT.

A few product feedback points from the portal assistant test:

• Spot holdings are not treated like tracked positions. After a fresh assistant session, I had to manually restate SOL size, entry price, TP/SL, and exit plan before it could monitor P&L.

• Longer assistant sessions repeatedly hit Prompt too long (max 32768 bytes) after several tool steps, even when my visible prompt was short. Starting a fresh session worked, so this may be hidden context/tool-output growth.

• A compact “monitor mode” would help: bounded reads, short responses, current wallet holdings, average spot entry, realized/unrealized P&L, user exit plan, and HOLD/EXIT status.

• Final fill reporting needs one authoritative source. After my SOL sell, the assistant first reported a much worse fill from the closedOrder row, then corrected itself using the history-canister fill record, which matched the quoted execution price. The trade itself appears to have worked, but the assistant should consistently use the settled fill/trade record for execution price and P&L.

• The fuel warning appeared and later disappeared. It would help if fuel/cycles status were available through a clear read endpoint or surfaced in a way agents can verify.

Overall, the built-in assistant is promising and already usable for small confirmation-gated spot workflows. The main gaps I saw are persistence, compact monitoring, clear spot trade journaling, and consistent settled-fill reporting.

The remaining open question from my prior post is still external MCP: should external ChatGPT MCP clients have access to the same safe proposal/preview/action surface as the portal assistant, or is the portal assistant currently using a different internal path?