notifications

name

notifications

description

Use when a QuantConnect/LEAN live algorithm sends data out or receives external instructions. Triggers — code uses py`notify.email`cs`Notify.Email`/py`notify.sms`cs`Notify.Sms`/py`notify.telegram`cs`Notify.Telegram`/py`notify.web`cs`Notify.Web`/py`notify.ftp`cs`Notify.Ftp`/py`notify.sftp`cs`Notify.Sftp`, py`signal_export.*`cs`SignalExport.*`, py`on_command`cs`OnCommand`, py`add_command`cs`AddCommand`, py`broadcast_command`cs`BroadcastCommand`, py`link()`cs`Link()`, or a class implementing `ISignalExportTarget` / `Command`; phrases like "Discord/Slack alert on fills", "email on drawdown", "push portfolio targets to our endpoint", "custom signal export to my broker", "manual liquidate from outside", "parent algo signals child", "multi-algorithm arbitrage". Skip when — purely in-algo log/debug (use `logging` skill).

External Communication in QuantConnect / LEAN

A live-trading algorithm has several primitives for talking to the outside world. The right choice depends on direction (out / in) and content (trade data / human alert / file / instruction). Two recurring decisions are easy to get wrong:

• Trade data going out: prefer Signal Exports over notify.web.
• Instructions coming in: prefer Commands over polling the Object Store.

Decision matrix

Direction	Content	Primitive
Out	Portfolio targets / trade signals to a platform	Signal Exports (custom ISignalExportTarget)
Out	Human-facing alert (fill, error, daily summary)	notify.email / sms / telegram / web
Out	Bulk file delivery to a counterparty	notify.ftp / notify.sftp
In	Manual instruction (liquidate, parameter change)	Commands (pyon_commandcsOnCommand or Command subclass)
In	Cross-algorithm coordination (parent → child, arbitrage)	Commands + project_id link, or broadcast_command

Critical: live-trading gating

notify.* and command-send primitives are Cloud live trading only — no-ops in backtests, local LEAN, and LEAN CLI live deployments. Signal Exports are different: they fire anywhere the algorithm runs, including backtests.

• For notify.* and command sends (pylinkcsLink, pybroadcast_commandcsBroadcastCommand, pydownload_datacsDownloadData): gate every call site with pyself.live_modecsLiveMode.
• For Signal Exports to a production endpoint: either skip pyadd_signal_export_providercsAddSignalExportProvider when not in live mode, or short-circuit at the top of send / Send.

Sending OUT

Trade signals / portfolio targets → Signal Exports

For portfolio targets to a fund, allocator, or trading platform: use Signal Exports, not notify.web. They plug into the trading workflow — a send call fires automatically on every fill with the current portfolio state — and the manager:

• Debounces fills into one batched call (default 5s window).
• Passes a standardized PortfolioTarget list, not an ad-hoc payload.
• Survives broker reconnects without duplicating the payload.

Implementation

Implement ISignalExportTarget — a class with pysend(parameters: SignalExportTargetParameters) -> boolcsbool Send(SignalExportTargetParameters parameters) and pydispose() -> Nonecsvoid Dispose(). Register in initialize via pysignal_export.add_signal_export_provider(...)csSignalExport.AddSignalExportProvider(...). pyparameters.algorithmcsparameters.Algorithm is the algorithm; pyparameters.targetscsparameters.Targets is the list of PortfolioTargets.

The only real question: what payload does the receiver expect?

QuantConnect doesn't define the wire format — the receiving service does. Before writing the class, get from the user (or the receiver's API docs):

1. Endpoint URL and auth. Bearer token, API key header, HMAC body? Lift to module-level constants or constructor args, not hard-coded inside send.
2. Field names and types. {symbol, quantity}? {ticker, weight}? {asset, side, size}? The receiver's contract drives every line of the body-build code.
3. Per-target or per-batch? Most APIs accept a list — send is called once per aggregated batch of fills, so build one request, not N.
4. Quantity in shares or in weight? pyparameters.targets[i].quantitycsparameters.Targets[i].Quantity is a portfolio weight (a fraction like 0.10), not a share count. If the receiver wants share counts, round-trip via pyPortfolioTarget.percent(parameters.algorithm, x.symbol, x.quantity, x.tag)csPortfolioTarget.Percent(parameters.Algorithm, x.Symbol, x.Quantity, x.Tag) — the returned target's quantity is the share count.

Pass tag through — it's the only carrier of caller-provided context (signal id, regime, confidence) and is empty unless the calling code created targets with a tag.

Three things that bite

• Reuse one HTTP client. Allocate pySessioncsHttpClient once in the constructor, close in dispose. Building a fresh client per send exhausts ephemeral ports under load.
• Return a bool, don't raise. Catch network and JSON errors and return False. An exception escaping send aborts the algorithm.
• send fires in backtests too. Either skip add_signal_export_provider when not in live mode, or short-circuit at the top of send with a live_mode check.

Human-facing alerts → notify.*

Six channels for messages a human (not a trading system) will read:

notify.email(address, subject, message, data=None, headers=None)
notify.sms(phone_number, message)
notify.telegram(id, message, token=None)
notify.web(address, data, headers=None)
notify.ftp(hostname, username, password, file_path, file_content, port=None)
notify.sftp(hostname, username, password|private_key + private_key_passphrase, file_path, file_content, port=None)

Channel	Body limit	Per-channel notes
Email	10 KB body, optional named attachment	headers={'filename': '...'} to override default attachment.txt
SMS	1,600 chars; E.164 phone (+1...)	Only channel with no free quota; per-message QCC cost (1 US/CA, 10 international)
Telegram	Plain text via bot in a group	Group ID is a negative integer; emojis must be UTF-32 escapes ('\U0001f680'); token is optional only if @quantconnect_notifications_bot is in the group
Web (HTTP POST)	300s receiver timeout	No retry. Discord receivers expect a content-keyed JSON envelope, not a plain string
FTP / SFTP	Sends file content under file_path	SFTP accepts password= OR private_key= (with optional passphrase), not both. FTP is plaintext

Hourly free quota by tier

Tier	Notifications per hour
Free	N/A
Quant Researcher	20
Team	60
Trading Firm	240
Institution	3,600

The hourly quota covers email, FTP, Telegram, and webhook combined; overage costs 1 QCC per notification. SMS is excluded and always billed per message regardless of tier.

Wrap repeated calls in a helper

When the same channel fires from more than one place, route every call through a private helper (e.g. _notify_email, _notify_sms) so the recipient/credentials and the live_mode gate live in one place. Lift recipients to module-level constants. One-off notifications can stay inline.

Fire on events, not per-bar

Putting notify.* inside pyon_datacsOnData on minute or tick resolution saturates the hourly quota in seconds and burns QCC. Notifications belong on events:

• pyon_order_eventcsOnOrderEvent — filled / canceled / rejected.
• pyon_brokerage_messagecsOnBrokerageMessage — broker errors, disconnects.
• pyon_warmup_finishedcsOnWarmupFinished.
• Scheduled events (daily summary, threshold alerts) — see the scheduled-events skill.
• Signal / state transitions — only on the change, not every bar the condition is true.

Send derived values only

The notification system can't be used for data distribution — that's a terms-of-use violation, not a quota concern. Send derived information (signal value, portfolio value, fill summary), not raw bars/quotes/trades from QuantConnect-subscribed datasets.

Receiving IN — Commands

Commands inject data into a running live algorithm. Typical uses: a manual "panic close-all", a click-to-confirm grey-box trade, or coordinating sibling algorithms (arbitrage between exchanges, child strategies offloading execution to a brokerage-connected parent).

Commands give you authenticated delivery with real-time pickup — no upload-to-Object-Store-and-poll round-trip — and both REST/API broadcast and project-targeted send (via project_id + pylink().download_data()csLink().DownloadData()) work for parent/child coordination.

Before writing any code: ask the user about the payload

Handler signature, Command subclass fields, and the sender script all depend on the payload shape. Before generating code, ask:

1. What fields will the command carry? (e.g. ticker + quantity, or action + target_weight.) Use those exact names — they become attribute accesses inside the handler.
2. One command shape, or several? One → generic pyon_command(data)csOnCommand(dynamic data). Several → encapsulated Command subclasses dispatched by $type.
3. How will it be sent? REST API, email click-link, broadcast from a sibling, or LEAN CLI. This decides whether to add a live_mode-guarded pylink(...)csLink(...) / pybroadcast_command(...)csBroadcastCommand(...) inside the algorithm.

Always link the user to the sender-script docs

Don't inline auth/REST boilerplate in the algorithm. Add a comment with a pointer to https://www.quantconnect.com/docs/v2/writing-algorithms/live-trading/commands#06-Send-Commands-by-API (or #07-Broadcast-Commands-by-API for org-wide broadcasts), plus the payload shape this handler reads.

Two handler styles: generic vs encapsulated

The dispatcher routes on whether the payload contains a $type key.

• Generic (single command shape): override pyon_command(data)csOnCommand(dynamic data). Payload without $type arrives here. Keys become attributes (pydata.ticker / data["ticker"]csdata.Ticker).
• Encapsulated (multiple distinct shapes): subclass Command (Python: class with class-level attributes; C#: properties), register with pyadd_command(MyCommand)csAddCommand<MyCommand>(). Payload with $type → matching registered Command subclass; other keys populate the instance fields. With $type set but the class not registered on the receiver, the command is dropped silently — relevant for broadcast_command.

Subclass-method access inside run / Run. algorithm is typed as IAlgorithm. Methods on base QCAlgorithm (log, set_holdings, portfolio, …) work, but methods you defined on your own subclass aren't reachable through it.

• Python: set MyCommand.ALGORITHM = self in initialize and call through the static (e.g. MyCommand.ALGORITHM.do_something()).
• C#: cast — ((MyAlgorithm)algorithm).DoSomething().

The return value (pyOptional[bool]csbool?) surfaces in the live commands log and in pydownload_data()csDownloadData()'s response body — always return True on success or False on rejection rather than implicit None.

Sending commands: four mechanisms

Mechanism	Reaches	Use it for
pyself.link(...)csLink(...)	One project (caller's by default; another via project_id) when the URL is opened	Email/Slack click-to-run, grey-box confirmations
pyself.broadcast_command(...)csBroadcastCommand(...)	Every live deployment in the org except the caller, except algos without the $type registered	Sibling-algorithm coordination
REST POST /live/commands/create	One project, by projectId	External orchestration (script, webhook, dashboard)
REST POST /live/commands/broadcast	Whole organization (with optional excludeProjectId)	External fan-out

LEAN CLI wraps the same REST endpoints. The receiver doesn't know or care which mechanism delivered the payload — all four funnel into the same pyon_commandcsOnCommand / pyCommand.runcsCommand.Run dispatcher.

The OrderCommand $type is built-in — every live deployment understands it without add_command and places the order described by the payload (symbol, order_type, quantity, limit_price, stop_price, tag).

Cross-project send for parent/child

Set pyalgorithm.project_idcsalgorithm.ProjectId to the target project before calling link. pyExtensions.download_data(link)cslink.DownloadData() executes synchronously and returns the response body — use it for the parent-child execution-offload pattern (paper-trading child sends commands to a brokerage-connected parent that places the actual orders).

Broadcast scope rules

• The sender does not receive its own broadcast. If you need fan-out plus local execution, run the local path first, then broadcast.
• A typed command is delivered only to algorithms that registered the same Command class via add_command; others drop it silently.
• A generic command (no $type) reaches every live algorithm in the org that defines on_command.
• LEAN does not add a sender field — always include py"sender": self.project_idcssender = ProjectId so receivers can disambiguate which sibling sent it.

Multi-algorithm coordination

Arbitrage and parent-child execution are the same pattern with different mechanisms:

• Symmetric N-to-N (e.g. arbitrage between exchanges): broadcast_command — each algorithm broadcasts fills from pyon_order_eventcsOnOrderEvent (gated on live_mode and pyOrderStatus.FILLEDcsOrderStatus.Filled) and reacts to siblings' broadcasts in on_command.
• One-to-one with response (e.g. paper-trading child offloading execution to a brokerage-connected parent): link + project_id + pydownload_data(link)cslink.DownloadData() — the child sets project_id to the parent and calls download_data from on_order_event to fire synchronously and read the parent's return value.

For a parent that just translates payloads into orders, set pyself.settings.seed_initial_prices = TruecsSettings.SeedInitialPrices = true in initialize so newly-added symbols can be traded on the first command.

Command gotchas

• LEAN uppercases the first character of every payload key when delivering to C#. A payload sent as {"ticker": "AAPL"} reads as data.Ticker in C# but data.ticker in Python. When defining an encapsulated Command subclass, name C# properties PascalCase (Ticker, Quantity).
• dynamic defers field-resolution to runtime. data.Ticker compiles even if the field doesn't exist; missing fields throw at runtime. Validate before use.
• Cast numeric fields explicitly — var qty = (decimal)data.Quantity; — JSON deserializes integers as int/long and arithmetic mixing types throws.
• Link(new { ... }) uses an anonymous object whose property names go on the wire verbatim. new { Ticker = "AAPL" } produces {"Ticker": "AAPL"}, which a Python sibling reads as data.Ticker (not data.ticker). For lowercase keys, use Dictionary<string, object>.