29 Canopus Assistant

Everything else in this handbook is free and open source. Canopus Assistant is an optional AI companion that rides inside the Polaris web interface. It comes in three flavours: one paid cloud tier, whose subscription revenue funds the ongoing development of Polaris itself, and two free tiers that run entirely on your own hardware. All three are strictly opt-in. Polaris ships with the assistant switched off, never depends on it, and works exactly the same whether or not you ever turn it on. This chapter explains what it does, how to choose and enable a backend, and, just as importantly, what it does not do with your data or your telescope.

Think of it as a co-pilot rather than an autopilot. It plans the night, proposes the actions to get there, answers astrophotography questions in plain language, and (on the cloud tier) inspects your frames. Anything that moves hardware or changes a running session waits for your approval first.

29.1 What it can do

Every backend shares the same core jobs. The cloud tier adds a few that need either a vision-capable model or more horsepower than a small local model has; those are called out below.

It plans the night. Tell it what you feel like shooting, “a nebula that’s up now”, “something for a two-hour window”, “narrowband from a Bortle 7 backyard”, and it ranks well-placed deep-sky targets (plus the Moon and planets), checks their altitude across the night, the weather and the Moon’s interference, and then assembles a ready-to-run plan in the night planner (Chapter 16) with sensible exposures, dithering, and refocus points. You review and tweak it before anything runs. (Building the full multi-target plan is a cloud-tier feature; the local tiers still suggest targets, exposures and framing.)

It drives the rig, with your approval. Slewing to a target, autofocusing (Chapter 11), starting a sequence or a live stack, and dithering are each proposed as an action you approve, review, or reject. When it acts, it navigates the Polaris UI to the relevant panel so you can watch focus, guiding (Chapter 12), or the stack as it works. (Parking the mount and stopping guiding or a live stack are cloud-tier extras.)

It knows astrophotography. Grounded, practical answers on focusing, guiding, exposure and histograms, filters, polar alignment (Chapter 9), calibration frames, and diagnosing bad subs, drawing on a knowledge base that includes this very handbook. Available on every backend.

It looks at your images (cloud tier). Ask “is this frame any good?” and it inspects the current frame, focus and star shape, trailing, gradients, framing, and tells you what to fix. It can also grade a folder of subs and drive the STUDIO stacking and post-processing tools (Chapter 20) on your behalf, again as reviewable steps. Frame analysis needs a vision-capable model, so it runs on the cloud tier only; the two local tiers are text only.

29.2 Choosing a backend

You pick where the assistant runs in Settings, Assistant. There are three choices, and the picker labels each by what runs it:

Cloud (subscription). The most capable option, and the only one that can look at your frames. The AI runs on our servers (gpt-5.3-chat on Microsoft Azure); your browser relays approved actions to your local Polaris. It needs an internet connection and a subscription (US$4.99 per month at the time of writing, with a short free trial). Any device that can open the Polaris web interface can use it, no local model or GPU required.
On this host (free). A small language model runs on the Polaris host itself, the machine Polaris is running on, so any phone or tablet on your network becomes a thin client. No account, no subscription, fully offline, text only. It needs a 64-bit host with roughly 6 GB of RAM free alongside Polaris; 8 GB or more and a faster processor let warm replies land in a few seconds (the first reply of a session takes longer while the model loads). Polaris downloads the model and runtime for you on first use.
On this device (free). The assistant runs the model on the machine you are browsing from, using a local LLM server you already run (Ollama, LM Studio, or llama.cpp), or, inside the Polaris mobile app, a model bundled into the app. This is the fastest option on a powerful PC or Mac with a capable GPU, where replies are near-instant. No account, fully offline, text only. When Polaris and your local LLM are on the same machine, Polaris routes the calls for you, so there is nothing to configure; only a split setup, where you browse a Polaris host on a different machine, needs the LLM server to allow the Polaris page’s origin.

The two local tiers are free and keep every message on your own hardware. The cloud tier is the only paid one, and the only one whose revenue funds Polaris.

Note

On a phone or tablet browser, “On this device” has no local model to run, because the browser is not the machine hosting the model. Use “On this host” instead (the model runs on the Polaris host and your phone is just a screen), or install the Polaris mobile app, which can run a small model natively on the phone.

29.3 Turning it on

Open Settings, Assistant and choose a backend.

For a free local backend, pick “On this host” or “On this device” and follow the short setup shown in the card: download the on-host model, or point Polaris at your local LLM and pick a model. The assistant’s floating button appears straight away; there is no account to create and nothing to pay.

For the cloud tier, the assistant also surfaces as a small, dismissible badge in the bottom-right corner of the interface. The flow from there is deliberately short:

Click the badge to open a short intro that explains what the assistant is and what it costs (and points you to the free local options, in case you would rather run it yourself).
Choose to subscribe. You create an account with just an email address, there is no password to invent. Polaris sends you a magic link (and a short numeric code you can type instead), and clicking it signs you in. The subscription itself is handled by Stripe’s secure checkout.
Once you are subscribed, a floating action button replaces the badge. Click it to open the chat panel and start talking to the assistant.

If you dismiss the badge, it stays gone until there is something new to show. Nothing about the cloud flow is baked into open-source Polaris: the app only knows how to fetch a small manifest from the assistant’s cloud and render whatever it describes. With no manifest configured, the badge never appears at all.

29.4 How it works

There are two architectures, depending on the backend, and both share the same safety and privacy model.

Cloud tier. Three parts fit together:

The cloud service holds the language model, the account and billing records, and the “agent loop” that turns your requests into a plan of tool calls. It is the only part that talks to the AI.
The chat panel is a small web app served from that cloud and shown inside Polaris as an embedded frame.
Your browser is the bridge. The cloud never contacts your telescope. When the agent decides to read the system status or slew the mount, it does not reach into your network; it hands that intent to your browser, which executes it against your local Polaris over your LAN, using the session you are already logged into, and hands the result back.

This has a happy consequence: because the browser is the bridge, your Polaris host (Chapter 27) never needs its own internet connection. The single-board computer at the telescope can sit on an offline field network; as long as the browser you are holding has both internet (to reach the cloud) and LAN access (to reach Polaris), the assistant works.

Local tiers. There is no cloud and no server-side AI. The agent loop runs in your browser (for “On this device”) or on the Polaris host itself (for “On this host”), talking to a local model over a plain HTTP endpoint. Prompts and replies never leave your hardware. The same “browser is the bridge” model still applies for actions: proposed tool calls run against your local Polaris through the session you are logged into.

29.5 Staying in control

This applies to every backend. Read-only questions, “what’s up tonight?”, “how’s guiding?”, “is this sub sharp?”, run freely; they only look. Anything that moves hardware or changes a running session is different: slewing, autofocus, starting or stopping a sequence or live stack, parking, a meridian flip. Those are never performed silently. The assistant proposes an action, and it does nothing until you approve it. You can review the exact steps, approve them, or reject them outright. The approval gate is structural, not a matter of the model’s good behaviour, so it holds on the local tiers exactly as it does on the cloud. The assistant is a co-pilot with its hands visible on the yoke, not an autopilot you have to trust blindly.

29.6 Your account and subscription

This section applies to the cloud tier only; the two free local backends have no account and nothing to pay.

The cloud tier is a flat monthly subscription (US$4.99 per month at the time of writing; the current price is always shown on the subscribe screen), with a short free trial for new subscribers. Sign-in is by magic link, so there is no password to manage; you can change your email or resend the link from the sign-in screen. Billing and cancellation go through Stripe’s standard customer flow.

To keep the shared AI service fair and its costs predictable, each subscriber has a generous monthly usage allowance. In normal use you will never notice it; if you ever reach it, the assistant tells you plainly and picks up again at the start of the next cycle. Your current usage is shown in the assistant’s settings menu.

29.7 Privacy

On the local tiers, privacy is simple: nothing leaves your hardware. There is no account, no cloud, and no third-party AI provider. Your messages, the model, and every reply stay on the Polaris host or on the device you are browsing from.

On the cloud tier, the assistant is designed to need very little about you. It does not upload or store your images. When it “looks at a frame”, the analysis happens against your local Polaris, and only the result, sharpness, star shape, a gradient note, travels back as text. Likewise, when it needs to know how the session is going, your browser forwards a compact, on-demand snapshot of status (not your data) so the agent can reason about it. The account itself is just your email address plus the subscription state. The full, current privacy policy is linked from the intro screen and hosted alongside the service.

29.8 Why the cloud tier is a separate product

Polaris is licensed under the AGPLv3, which is what keeps it free and open, and the two local backends are part of that open codebase. The cloud tier is the one closed, commercial piece, and it is kept cleanly separable on purpose: it runs as an entirely separate service and only ever talks to Polaris through the same public web API that any other client uses. Nothing in open-source Polaris contains the cloud tier’s product logic; the open app is just a generic, off-by-default host that renders what the cloud manifest tells it to. That boundary is what lets a paid add-on fund a free application without compromising either one.