OpenClaw狂潮下，CEX爭搶AI Agent交易入口

Original | Odaily (@OdailyChina)

Author | Ethan (@ethanzhang_web3)

"Lobster Fever" has spread to crypto exchanges, becoming one of the most interesting scenes in the OpenClaw frenzy. In recent days, OKX, Binance, Gate, and Bitget have almost simultaneously pushed Skills, MCP, CLI, Agent Hub, and Wallet Skill into the spotlight.

This time, the exchanges are not just chasing a trend; they are transforming their capabilities, previously encapsulated within apps, web pages, and APIs, into an interface layer that Agents can understand, call upon, and execute. In the past, the competition was about liquidity, fees, and listing speed; now, they are starting to compete for a more front-end position: when users delegate research, filtering, and pre-order preparation to an Agent, which platform will be the first to be called upon.

This is also the truly crucial aspect of this round of changes.

First, let's look at what these exchanges have been updating in recent days

In this round of action, the update pace of each exchange has been very dense.

OKX's moves actually unfolded in two steps. On March 3rd, Odaily cited official news stating that OKX OnchainOS had opened AI-related capabilities, allowing Agents to execute on-chain transactions, gain market insights, perform address analysis, and make on-chain payments through three methods: AI Skills, MCP, and Open API. According to the OnchainOS official documentation, this system itself already covers core capabilities such as wallet, trading, market data, and payments, with documentation claiming coverage of 130+ mainstream public chains and aggregation of 500+ DEX routes on the trading side. By March 10th, the OKX Agent Trade Kit took another step forward, bringing the two access paths, okx-trade-mcp and okx-trade-cli, along with 4 pluggable Skills, to the forefront. What's truly noteworthy is not the "AI trading" label, but the detailed explanation: AI can complete spot, perpetual, and conditional order operations through a single MCP interface; keys are only stored in local configurations, with signing done locally; if an API Key lacks trading permissions, the corresponding order tool simply won't be registered in the toolbox. The official scale given is 82 tools across 7 modules, also mentioning live options market and simulation mode. Looking at OnchainOS and Agent Trade Kit together, it's clear OKX isn't just aiming to build a trading assistant this round, but is aligning both on-chain and trading capabilities towards Agent infrastructure.

Binance's pace started earlier and has been pushed forward continuously. On March 3rd, the first batch of 7 AI Agent Skills were released, followed by the Skills Hub bringing these capabilities to the forefront, publicly listing modules like crypto-market-rank, meme-rush, query-address-info, query-token-audit, query-token-info, spot, and trading-signal. The next day, it was disclosed that MPC Wallet and DeFi Wealth Management had entered the audit stage; by March 12th, 4 new AI Agent Skills directly expanded capabilities to Alpha market data, USDⓈ-M futures trading, margin trading, and asset management. This means Binance's line is no longer just the initial 7 informational Skills, but has rapidly extended from an information gateway to derivatives, leverage, and account management. The most interesting line on the page isn't the feature introduction, but that all Skills undergo review before going live. This indicates Binance wants to capture not just the traffic gateway, but also the rules for capability distribution in the Agent era.

Gate's updates feel more like a rapid-fire sequence. On March 5th, Blue Lobster was released first, lowering the barrier for ordinary users to experience GateClaw; on March 7th, DEX MCP went live; on March 10th, CEX MCP followed, by which point Gate had simultaneously connected both on-chain and centralized trading capabilities to the Agent scenario. Then, on March 11th alone, it consecutively pushed forward the Skills Hub, new version of Blue Lobster, Gate CLI, Blue Lobster Operation Guide, and 20 AI Agent Skill updates. This means Gate isn't just putting up a single "Gate for AI" page; within days, it has laid out the experience gateway, DEX/CEX MCP, CLI, Skills Hub, and Agent Skill layers. Looking back at the architecture of Gate for AI—application layer, capability layer, protocol layer, infrastructure layer, and the five core modules of Exchange, DEX, Wallet, News, and Info—it becomes clearer: its goal isn't a single-point tool, but packaging CEX, DEX, wallet, information, and on-chain data together into a suite of Agent infrastructure.

Bitget's actions have also formed a very complete line. The earliest was on February 27th, with the Bitget Wallet Skill beta released first, focusing on allowing large models and automation tools to access on-chain data and trading infrastructure using natural language, with transactions still requiring user signature confirmation; then on March 2nd, Bitget Wallet initiated Agent scenario capability exploration and launched the Skills beta, bringing the Wallet line fully to the forefront. By March 9th, the Bitget Agent Hub received a major upgrade, connecting the Skills and CLI modules to form a complete invocation system with the MCP and API launched last month, with official claims of 9 modules, 58 tools, and 3-minute integration with OpenClaw; on the same day, Bitget Wallet announced integration with Paydify, bringing consumer payment scenarios into the Agent ecosystem; then on March 12th, Bitget Wallet MCP opened for user testing, further pushing on-chain wallet capabilities towards a callable interface layer. Looking further, the March monthly report presented another set of numbers: approximately $205.95 million net inflow in February, ranking third globally among CEXs, with BTC reserves rising to 36,700. Connecting these actions, Bitget is clearly not just temporarily riding the lobster wave; while experiencing platform growth, capital inflow, and brand expansion, it is integrating Wallet, payments, Agent Hub, MCP, and other capabilities into its infrastructure narrative.

Looking at the timeline of these past few days reveals a clear change: This wave is no longer just exchanges collectively issuing a round of AI press releases; several leading platforms have started, one after another, to repackage capabilities like market data, addresses, audits, wallets, order placement, and risk control—previously scattered across pages and APIs—into modules that Agents can call upon.

In a nutshell, the difference is: Previously, most products only made AI better at talking; this round, leading exchanges are starting to make AI capable of actually calling things.

The industry hasn't been silent about AI trading over the past year. Copy trading, signal bots, strategy generation, research report summaries—all have been discussed. The problem is, those products often just added a smarter front-end to the existing trading process. AI analyzes on one side, trading executes on the other, with users still needing to switch pages, copy parameters, and click confirm in between. Essentially, they were still helping you watch, not connecting the systems for you.

This round is different. It's no longer satisfied with keeping research and suggestions within a dialog box, but is starting to move towards system calls, permission boundaries, and execution chains. Precisely because of this, what exchanges are now releasing is not just a simple enhancement of the conversational layer, but starting to touch real system interfaces.

Comparing the players, the gap is no longer about "having it or not," but "how far it's been implemented"

Based on my personal hands-on experience, if we truly put the four into a comparison table, spot trading and conditional orders are no longer scarce capabilities. The gap mainly lies in deeper aspects. (Relevant Agent tutorials are attached at the end of the article; here I'll only share my impressions after testing.)

Starting with the most basic spot scenario. Buying some ETH while placing take-profit and stop-loss orders—this kind of action is already supported by three. Binance's Spot Skill supports OCO, OKX's spot module can handle take-profit/stop-loss, and Bitget's spot conditional orders are also available. At this stage, the difference isn't about capability, but about which integrates more smoothly and whose Agent interprets intent more accurately.

Futures start to create separation. OKX and Bitget can already directly handle instructions like opening positions, stop-loss, and take-profit. Binance did not put futures at the forefront during its first batch of Skills phase, so that version felt more like a research and spot execution gateway. Although it later added USDⓈ-M futures, margin, and asset management, judging from the current publicly available product completeness, the smoothest parts remain the information layer and standardized spot scenarios.

Looking further, Bitget's boundaries are more expansive. Modules like copy trading, wealth management, and account management are already publicly displayed. Trader screening, automatic copy trading activation, wealth product queries, and subscriptions—these are no longer just slogans. OKX and Binance haven't yet placed these parts at the same depth. Therefore, Bitget gives a more direct impression: it's not just making a few more Agent tools; it's moving the entire trading environment into the dialog box.

Binance also has its own strengths. Among several public walkthroughs, the chain of address query, hot token analysis, token audit, and spot trading is the smoothest. Especially in the pre-order layer, capabilities like wallet address insights, token security audits, and market rankings are well-suited for Agents to handle first. However, its boundaries are also clear; for example, wallet queries currently only support BSC, Base, and Solana chains. Many capabilities first establish the gateway, then gradually add depth.

OKX seems more focused on the execution layer. By putting spot, perpetuals, conditional orders, options, local signing, and simulation environment together, it's clearly prioritizing solving a harder problem: once an Agent truly interacts with the order system, how are permissions managed, risk controls enforced, and simulations run? OKX's thinking here is evidently more forward-looking.

Gate is not yet easily comparable using single-point scenarios. Compared to the previous three, its publicly visible third-party hands-on scenarios are still relatively few, making it hard to directly claim it surpasses anyone in a specific trading action. However, judging from the consecutive rollout of DEX/CEX MCP, CLI, Skills Hub, and 20 Agent Skills in recent days, Gate isn't patching a single feature; it's laying a foundational layer. In the short term, it might not be the most powerful in usage; in the medium term, it aims to capture a more critical platform position. Beyond features, there's another intuitive feeling: even in naming, exchanges are becoming increasingly similar in this wave. While heat rises, distinctiveness hasn't fully kept up.

GateClaw Official Website

Using Bybit as a control group makes the difference more apparent. As of March 13, 2026, its most prominent public actions remain event-driven narratives like the AI vs. Human 1v1 Trading Competition, which can drive traffic but are clearly not the same product rhythm as the others pushing Skills, MCP, CLI, and modular interfaces forward.

Therefore, looking at the players together, the conclusion is already quite clear: Binance first secured the information and Skill distribution gateway, OKX is closest to a trading execution loop, Bitget currently shows the deepest public business vertical, Gate is more like building a platform foundation, and Bybit remains at the activity and communication layer, not yet entering this round's real product competition.

Why are exchanges the ones rushing out first?

This question is actually more important than who is currently doing more. Exchanges are precisely the companies least likely to slow down in this matter.

Market data, depth, accounts, orders, wallets, risk control—these are inherently the most mature, structured, and modularizable set of capabilities within an exchange's daily operations. For large models, the difficulty has never been "understanding a human sentence," but whether there are reliable external systems to connect to after understanding. Exchanges happen to have all these ready-made systems in hand.

A more realistic layer is that exchanges fear losing the gateway more than other projects. In the past, users opened the app first, then checked the market, then placed orders; later, users started entering through wallets, quantitative tools, Telegram groups, and on-chain dashboards. Now, Agents have created a new layer of gateway. In the future, users might not open a trading page first, but instead say in Claude, OpenClaw, ChatGPT, or a terminal: "Help me see which coins have the biggest moves today, and if the risk is controllable, give me a batch buying plan." If the first touchpoint becomes a dialog box, exchanges that don't proactively make themselves the default capability layer for Agents to call can easily be relegated to mere liquidity backends.

The OpenClaw wave of heat has just pushed this forward. Previously, terms like Skills, MCP, and CLI were more developer-oriented; now exchanges, media, and KOLs can use them to tell stories. For exchanges, this isn't just a few new buzzwords, but the sudden formation of a new layer of distribution channels. Whoever enters first has the chance to secure a position before standards solidify. (Recommended reading: