Commercial Agentic Trading Products

Mid-2026, a handful of live commercial products put an LLM or AI agent in front of a real brokerage account — and they differ enormously in what “agent” actually means: a research assistant that never touches an order, an official connector that drafts instructions a human must approve, and automation that executes once you activate it. This page compares five of them on what they actually do, not their marketing copy.

The Spectrum: Research Assistant → Draft-Only → Full Execution

Ordering these five by how much autonomy the agent actually has reveals three distinct tiers, and they do not map cleanly onto how “AI-powered” each product's marketing sounds.

Tier 1 — Research and Insight Only, No Execution

Robinhood Cortex sits here. Despite “prompt-to-trade”-sounding marketing in some secondary coverage, Robinhood's own support documentation is explicit: Cortex “is not placing trades for you.” Digests explains why an asset or a portfolio moved, in plain language, synthesizing market data, news, and analyst reports. Trade Builder surfaces price signals, technicals, and news to screen matching options strategies for a market view the user states. The human enters every order.

Tier 2 — Reads and Drafts, Human Approves

Interactive Brokers' AI Integrations is the clearest example of the MCP-connector pattern this site's research-landscape coverage flags as an emerging category: an official, vendor-built connector (not a community hack) that lets Claude, ChatGPT, or Grok authenticate directly against a client's existing IBKR account through that AI platform's own connector directory, with no API key or password shared with the AI vendor. The assistant can read positions, analyze risk, and draft trade instructions across 170+ markets — equities, ETFs, options, futures, and futures options — but every draft lands in a dedicated “AI Instructions” tab inside IBKR's own platform, requiring manual client review and approval before anything reaches the market. Launched per a 2026-06-22 press release, Claude first, ChatGPT and Grok added in the same release.

Tier 3 — Executes Once Activated

Public.com's AI Agents land here: a user's natural-language prompt is refined into a reviewable, rule-based automated workflow, and once the user activates it, it runs on Public's own order-routing infrastructure, continuously monitoring real-time market data and firing the trade when the stated criteria are met — no external API key required. Public's page does not disclose whether the underlying mechanism re-invokes an LLM on every triggering event or compiles to a static rule at setup time, which matters for how “agentic” the running system really is versus a well-packaged conditional order. JPMorgan's LOXM is also full-execution, but institutional-only: a deep-reinforcement-learning engine, not an LLM, deployed inside JPMorgan's own European electronic-equities business since Q1 2017 to choose order size, price, and timing that minimizes market impact, operating under the firm's own internal risk framework rather than a client-facing product with a sign-up page.

Off to the Side — A Coding Tool, Not a Decisioning Tool

MetaTrader 5's AI Assistant (still commonly nicknamed “CoPilot” in trade press, though that was never MetaQuotes' own name) is the least agentic of the five in its stable form: a user writes a trading idea as a code comment, and the assistant generates matching MQL5, Python, or C++ code, or explains existing code. It does not decide when to trade. A newer beta, build 6030 (released 2026-07-16), moves toward the other four by adding native AI agents and MCP support so external agents can connect to charts, market data, accounts, and execution workflows — but that capability is beta, not yet on the stable channel, as of this page's verification.