A desktop app, a local bridge and rules you can read.

Connecting MetaTrader

1. Install MetaTrader 4 with MetaEditor from your broker and sign in to the account you intend to use. Start with a demo account. Windows and Apple Silicon Macs are the supported computers; Intel Macs are not supported.
2. Install the HttpBridge Expert Advisor into MetaTrader and compile it. The app offers a guided setup that installs and compiles the Expert Advisor, writes its inputs and starts MetaTrader with it on a chart.
3. Allow the bridge address in MetaTrader under Tools, Options, Expert Advisors. MetaTrader keeps that list encrypted, so this remains a manual step.
4. Enable AutoTrading. The app header shows whether the Expert Advisor is online.

MetaTrader only makes web requests to ports 80 and 443 and refuses the names localhost and 127.0.0.1. The bridge therefore listens on port 80 on your own computer under a hosts-file alias. Port 80 must be free; a web server bound to it must be stopped first.

The local bridge

The bridge accepts connections only from your own computer. The Expert Advisor polls it about once a second, authenticating with a per-install key that is generated for your installation and shown only during setup. Wrong keys are refused, repeated wrong keys lock the bridge briefly, and the right key always passes so the Expert Advisor cannot be locked out.

Commands follow an at-most-once lifecycle. A queued command the Expert Advisor does not collect within a short window expires and never reaches MetaTrader, so an order tapped while the Expert Advisor is offline does not fire minutes later. A dispatched command without an acknowledgement becomes unknown and is never re-sent, because that could double a trade; market orders carry a request reference so they can be reconciled against open positions.

The risk manager

Every trading command from any source, whether you, an agent or automatic mode, passes the same checks before it is sent. Refused commands are written to the audit log with the reason.

• Pre-checks without market data: the halt switch, a daily-loss halt, maximum lot size, a required stop loss, orders per day, trading sessions in local time and news blackouts by symbol or currency.
• A pre-trade snapshot for market orders: equity, open positions in total and per symbol, total lots, broker lot limits and the money at risk as a share of equity. If the snapshot fails, the order is blocked.
• Serialized trading commands, so two orders cannot both pass a limit that only one of them fits in.

The daily-loss limit takes the first equity reading of the trading day as its baseline and keeps it across restarts. At the limit, new orders halt until the next trading day and, if you chose it, open positions are flattened. Closing positions is always allowed, even while halted. Limits are yours to configure and can be switched off; they protect against the conditions you set, not against every loss.

Agents and providers

The AI side runs on a provider you configure and pay for separately: the Claude command-line tool with your own sign-in, the Codex command-line tool with your own sign-in, a local Ollama model, or a Claude API key stored in your operating-system keychain. The app does in code everything that code can do and calls the model for judgement: reading news, weighing fundamentals against technicals, writing briefs and answering your questions.

Agents have no trading tools. A lead agent can only propose; proposals are schema-checked, symbols validated, stops required and bounded, and the deterministic executor sizes and submits the order through the risk manager. News and feed text is treated as untrusted input: it is reduced to plain text, truncated and wrapped as data, and text that reads as instructions to the model is set aside in code before any agent sees it.

A usage ledger records each model call with its task, provider, tokens and estimated cost, so you can see what the AI is doing and what it costs under your own provider agreement.

Automatic mode

Automatic mode is off by default. When you enable it, the app executes only proposals that pass the risk manager and additional code checks at that moment: the technical score agrees with the direction, the spread is small relative to the stop, the target pays at least the risk, the trade does not hedge a held position, no high-impact release is due shortly, the book's open risk stays under your cap, exposure to the same currency is limited, no cool-off is running after consecutive losses and the proposal's signal quality meets your minimum.

Automatic mode trades only watched or held symbols and only ever tightens stops. A proposal that fails a check waits for you, with the reason. It is not supervision-free trading, and it does not change the risk of loss inherent in trading.

What works without AI

Trade alerts for positions near a stop or target, the economic calendar with automatic blackouts around high-impact releases, a measured outlook, a technical morning brief, indicators, a signal board that grades every watched symbol, the backtester with a trust analysis of its own results, the trade journal scored from closed trades, a book check across open positions, quick answers to routine questions and the risk manager all run in code without a provider.

Read what is stored on your computer and what leaves it in risk, data and security, then check platform access.