Frequently asked questions about NorthernGo

What can I build with NorthernGo?

You can build working web applications without writing code. In the cloud it suits CRMs, landing pages, dashboards, portfolio sites, internal tools and forms. The AI generates clean Vanilla JavaScript and Tailwind CSS, which compiles into a fast PWA from day one. It works best for self-contained apps built from scratch rather than large existing systems. Locally the models are much smaller. Stay with simple pages, forms and small changes, or read “What can the local models do”.

Why choose NorthernGo over other AI builders?

NorthernGo gives you standard code without locking you into a closed ecosystem. Most platforms push you into frameworks that are hard to move away from. NorthernGo generates plain Vanilla JavaScript and Tailwind CSS that you own entirely. We are also the only app builder we know of that runs generation locally on your own device via WebGPU, rather than only on someone else's servers.

Can I get more cloud generations on the free plan?

Yes. Free includes 5 cloud generations per calendar month. Share a saved app to the community — from the editor or from My Projects — and you get 5 more that same month, once. The app has to be a real page: a heading, a button or a form. An empty shell, or a script that does not run, is not published. Further shares that month still appear, without more generations. Taking the app down does not remove the generations you already received. Premium and Pro are not included. Those plans already have 25 and 50 cloud generations.

Which AI models does NorthernGo use to generate code?

In the cloud, NorthernGo runs on Gemini 3.8 Flash (Google) for code, architecture and image analysis. Locally in your browser via WebGPU (quantised, 3–9 billion parameters — not in the same class as the cloud model). All five are free on every plan, including Free. What decides which one you can run is free graphics memory; the figures below are rules of thumb for the download plus headroom, not a measurement of your machine: • Ministral 3 Reasoning (3B) — around 3–4 GB free. Smallest and fastest to load. Weakest of the set; use when memory is tight. • Qwen 3.5 (4B) — around 4 GB free. The default. For small, clearly described pages, forms and simple changes. Not for a first CRM or a multi-view app. • Qwen 2.5 Coder (7B) — around 6 GB free. Code-tuned. Often the most reliable of the local set on HTML and JavaScript. • Llama 3.1 (8B) — around 6 GB free. A general 8B model in the same class as the others — not a reasoning leap over the cloud. • Qwen 3.5 (9B) — around 7 GB free. The largest of the local models. Slightly better than 4B on longer instructions. Still far from the cloud model. On Pro you can also run Ollama on the same machine. The settings dialog has an Add a model section: a shortlist grouped by how much memory the computer needs (16, 32 or 60 GB), and a field where you paste a Hugging Face link. NorthernGo reads the model card, shows roughly how large the download is, and pulls it into Ollama with a progress bar — you do not have to use the terminal. Only GGUF text models work; embedding and vision-only repositories are rejected. Anything imported that way is marked untested in the builder: it will run, but the build pipeline has only been measured against the shortlist. The local models are open weights from Alibaba, Meta and Mistral. NorthernGo does not train its own models — what we build around them is the patch engine and in-browser syntax checks. There is no LLM critic rewrite on the local path.

What can the local models do, and what can they not do?

Worth being straight about, because the gap is real. A quantised model with 3–9 billion parameters on your own GPU is not comparable to Gemini in the cloud. That gap does not disappear because the engine around it is good. The default on Free and Premium is Qwen 3.5 4B — the model most people will actually run. Works well locally: • Self-contained apps from scratch — landing pages, forms, simple pages and small tools • HTML, CSS, Tailwind and plain JavaScript • Boilerplate and standard patterns • Small, clearly described changes to an app you already have (copy, colour, an extra field) Works poorly or not at all locally: • A first draft of a CRM, shop or dashboard with several views — the 4B model often produces a polished first screen and an unfinished third • Apps that do not fit the model's fixed context window (often 4–8 thousand tokens on WebGPU) • Deeper debugging of bugs that are not visible directly in the code • Architecture and state that must stay in sync across several views • Long, compound instructions with many conditions at once Our recommendation: do the first structural pass in the cloud. Then iterate locally on bounded changes — unlimited runs, zero cost per generation, and the prompt stays on the machine. Switch back to the cloud for the next genuinely structural step. You can switch mid-project.

What hardware do I need to run the AI locally?

You need a browser with WebGPU and enough free video memory (or unified memory) for the model you pick. NorthernGo checks that WebGPU exists. It does not measure VRAM or pick a model for you — if the weights do not fit, loading fails. Rule of thumb for the quantised models we ship: • Ministral 3 Reasoning (3B) — around 3–4 GB free • Qwen 3.5 (4B) — around 4 GB free • Qwen 2.5 Coder (7B) and Llama 3.1 (8B) — around 6 GB free • Qwen 3.5 (9B) — around 7 GB free The 4B model often runs on a current laptop, including Apple Silicon, via unified memory. The 7–9B models usually want a dedicated GPU with several gigabytes free; on integrated graphics they may refuse to load or run slowly enough that a cloud round trip would have been faster. On a phone, local AI is blocked only when WebGPU is missing or the browser reports under 8 GB. A reported 8 means “8 GB or more” — the page cannot see 16. Cloud builds work on a phone either way. If loading fails, use a smaller model or switch to the cloud.

How large is the model download, and does it happen every time?

The model weights are downloaded once and then cached in the browser. The next time you build, the model starts from cache with no download. Size depends on the model: about 2.2 GB for Ministral 3B up to about 5.5 GB for Qwen 3.5 9B. The first download therefore takes a while on a slow connection. After that, generation itself can run without a network. Saving a project, publishing, and the database in a generated app still need the internet. If you clear browser storage the cache is lost and the model is downloaded again.

Which browsers support local generation?

Local generation is built on WebGPU. In practice that means current Chrome, Edge and other Chromium browsers, including on a phone that has WebGPU. Safari on recent macOS can work. Firefox support varies by version and platform. On a phone the same WebGPU check applies. Safari often does not report memory; Chrome on Android reports at most 8 GB. Under 8 GB, local AI is blocked. A reported 8 means 8 GB or more. Cloud builds work on a phone either way. If the browser has no WebGPU, NorthernGo says so and you use cloud mode instead. Everything else in the platform works either way.

Who owns the code?

You own the generated code 100%. NorthernGo is not a locked-in platform. We generate raw, standard code. Copy source in the editor is Pro only. Premium and Pro users can download the whole project as a standalone ZIP, and Pro users can also push it directly to a GitHub repository — either way, you host it wherever you want. The prompts and the context sent with a cloud build belong to you as well. They are not used to train Google's models. See “Does Google train on my prompts, context or the code NorthernGo generates?”.

Am I locked into NorthernGo, or can I export my app?

You are never locked in. Premium and Pro users can export their entire app as a standard ZIP file containing plain Vanilla JavaScript and Tailwind CSS. Pro users can additionally push the project straight to a GitHub repository. The exported code is fully standalone, automatically configured with your project ID, and ready to be hosted anywhere.

Are the generated apps responsive, and do they work on mobile?

Yes, apps generated with NorthernGo are responsive and mobile-friendly from the start. Built with Tailwind CSS, the layouts adapt to smartphones, tablets and desktops. To enable full Progressive Web App (PWA) features and publish live, you first create and configure the subdomain in the project settings.

Can I generate an app from an image or sketch?

Yes, Pro users can build source code directly from sketches or reference images. Upload a picture or a design layout in the prompt area and the AI agent team analyses the image and generates a matching layout and functionality. Image analysis runs in the cloud, not locally.

Can I upload documents like PDFs or Word files to guide the AI?

Yes, you can upload project requirements directly. Pro users can upload PDFs, Word files (.docx) or plain text files (up to 10 MB). The AI agent team analyses the text to understand your requirements and builds your system based on that documentation. Document analysis runs in the cloud, not locally.

Is my data secure, and is NorthernGo GDPR compliant?

Data security and privacy are fundamental design choices in NorthernGo, and one distinction is worth being clear about. With Local AI (WebGPU), the local model processes your prompt and writes the code on your own GPU. On Premium or Pro, a new local build may still send a short structure request to the cloud Architect when you are online — that step is skipped offline, on Free, and on updates. The text you type in the prompt box before you have an account is stored only in this browser, so the idea survives registration. That is functional storage, not analytics, and it is not sent to our servers until you generate while signed in. Empty the box, start the build, or clear site data to remove it. As soon as you save a project, publish an app to a subdomain, or use the database, authentication or email features in a generated app, that data goes to our servers, because that is what those features do. For cloud workflows, all sensitive data, connections and keys are encrypted at rest. Cloud generation, on every plan including Free, uses Google's commercial Gemini API with billing on. Your prompts, context and generated code are not used to train Google's models. See “Does Google train on my prompts, context or the code NorthernGo generates?”. You can wipe all collected data, reset keys, or delete your account and everything tied to it at any time from the profile menu.

Does Google train on my prompts, context or the code NorthernGo generates?

No. Cloud app building goes through Google's commercial Gemini API on a Cloud project with billing enabled. That is true on every NorthernGo plan, including Free: the call uses our billed project, not Google's unpaid quota. Your prompts, the context sent with the build (system instructions, existing code and uploaded files) and the generated code belong to you. Content your end users write in a published app belongs to them and to you as controller. Google does not use that material to train or improve its models. On Google's unpaid API, outside the EU/EEA, Switzerland and the United Kingdom, staff may review disconnected anonymous samples to improve the products. That human-review stream is closed when billing is on, which is how NorthernGo calls Gemini. App building runs in an isolated, commercial API environment that follows the enterprise standard for data protection, under Google's data processing addendum for products where Google is a processor. Google may still keep prompts and responses for a limited time solely to detect violations of its use policy and to meet legal duties. That is abuse monitoring, not training and not the manual product review. Local text generation (WebGPU or Ollama) does not leave your device. Image generation, document upload and cloud mode still do. If you connect your own API key (Claude or GPT), that provider's terms apply, not Google's paid terms.

Do generated apps include GDPR deletion and data export by default?

Yes. GDPR Article 17 (erasure) and Article 20 (portability) are wired into the app runtime rather than left as something you remember to build. That holds for cloud Generate, local Ollama or WebGPU, MCP create_app / generate_app, and HTML import. With login, signed-in users get visible Download my data and Delete my account. Without login, apps that collect an e-mail get a quiet Privacy / Integritet control in the corner — games and calculators do not. It opens the same overlay: enter the e-mail that is on the row, and we send a confirmation link — one for a JSON download, one for erasure. Delete hides matching rows immediately; after 14 days they are permanently removed. Rows with no e-mail cannot be tied to the person. You remain the controller. An EU database and a working delete path are architecture, not a certificate. If you connect your own database, cascade deletion applies to NorthernGo's default store; you still have to delete what you routed elsewhere. To hide the Privacy control, put data-ng-hide-legal on the body — then you must offer another way to reach the rights. We do not know another AI app builder that ships this as the default in generated apps rather than as an Enterprise add-on or a paragraph in a policy.

Are generated apps security-checked before I see them?

Yes. Every landed app — cloud Generate, local Ollama or WebGPU, MCP create_app / save_app, and HTML import — runs through a deterministic scan of the source and of the mounted document. It redacts leaked keys, disables eval / new Function / document.write, neutralizes javascript: URLs, and drops script and iframe sources that are not on the allow-list (Tailwind, cdnjs, and the embeds we already ask the model for: YouTube, Vimeo, Spotify, OpenStreetMap, Google, Stripe). That is a guard, not a pentest. It does not prove the app is free of logic bugs or XSS via innerHTML. CSP on the published subdomain and the sandboxed preview iframe are the other layers. You still own unread flaws in the code, the same as if you had typed it.

How does NorthernGo update my app without breaking the existing design?

NorthernGo uses SmartSync, driven by an automatic app manifest. When you request a change, the engine reads a stored structural map of your existing components, tables and colour tokens, then applies a targeted patch instead of regenerating the whole app. That means your existing design and logic normally stay intact, and small changes stay small. Large or vaguely described changes are still more likely to affect more than you intended, so describe them one at a time. On a computer you can also point at an element in the preview instead of describing it from memory — see “Can I point at something in the app and ask for a change”.

Can I point at something in the app and ask for a change?

Yes, on a computer, when the next change will go to the cloud (Gemini). When the preview is ready, the Point button sits next to Save and the others above the app. Press it, click the thing you want to change, type what should happen and press Change. Escape cancels. It uses the same path as describing the change in the chat — the engine is told which element you pointed at. Small, specific requests (copy, colour, an extra field) work best. Large or vague instructions can still touch more than you meant. Point is not offered for local engines (WebGPU or Ollama). They are much weaker at targeted patches. Switch to the cloud engine, or write the change in the chat. There is no Point button on a phone.

What happens if the AI generates code with bugs or errors?

On the cloud path, every generation goes through an automated quality check before it reaches your screen. If the system detects missing tags, malformed markup or broken scripts, the Critic can attempt a repair before you see it. That catches a share of structural errors, not logic errors. On the local path there is no LLM rewrite. A syntax check blocks JavaScript that will not parse — a new build is not opened, and a patch that would break the app is rejected so the previous version stays. It does not attach missing click handlers or fix the wrong total. AI-generated code can still contain logic that does not do what you meant, especially on the smaller local models. Test the app before you publish it, and describe what went wrong in plain language.

How does NorthernGo ensure that the generated JavaScript code actually works?

On the cloud path, before a code change is delivered, NorthernGo runs a server-side syntax check: scripts are compiled in an isolated virtual machine. Code that will not parse is not shown as a finished app. On the local path the same idea runs in the browser: a syntax check blocks JavaScript that will not parse. It does not rewrite the app. To be precise: both steps catch syntax errors, not logic errors. A button with no handler attached is valid JavaScript and passes. That is why testing before you publish still matters.

Can I build apps that include built-in AI features like Text-to-Speech or image generation?

Yes, Premium and Pro users can embed AI capabilities directly into their own apps. NorthernGo gives your applications access to text intelligence, image generation, speech-to-text transcription and Text-to-Speech through simple global hooks. These can run either through our cloud proxies or offline in the browser via local hardware execution, depending on the capability.

What database does NorthernGo use, and can I connect my own?

Every app generated with NorthernGo comes with a built-in, automatic Supabase integration for your database and user authentication. You do not need to configure anything to get started. If you need a custom setup, Pro users can connect their own external database, a remote MCP server, Bokio (see “How do I connect Bokio to my app”), Gmail (see “How do I connect Gmail to my app”), or route data through webhooks (such as Zapier or Make) directly from the project menu. A Gemini- or MCP-created Pro app can also optionally expose those same collections as a hosted MCP server — see “Can my generated app become an MCP server that others connect to”.

Where is my app hosted, and do I need to configure Git or servers?

No, you never have to manage servers, terminal commands or Git to publish. A saved app is already served over HTTPS at its live link. Premium and Pro publish an installable PWA on a .northerngo.com subdomain from My Projects → More → Create PWA.

How do I deploy and launch my app to the cloud?

When the build finishes, the app is saved and already served over HTTPS, connected to its database. In My Projects, Live opens that address and Copy copies the link. Publish in the account menu shows the same link. A name of your own, as an installable PWA, is My Projects → More → Create PWA (Premium and Pro). On Free the address stays northerngo.com/?app=…

How do I connect my own custom domain name?

Pro plan users can link their own domains. Step 1: Open My Projects, click More on your application and select Custom Domain. Step 2: Enter your domain name and save. Step 3: At your DNS provider (Loopia, One.com and so on), create a CNAME pointing the domain at the target address NorthernGo gave you. Step 4: Add the same domain under Custom Domains on the Render service. SSL is issued after that step, usually within an hour. A CNAME alone does not issue the certificate.

Can I use a custom domain, or does it have to be a PWA?

Both, depending on your needs. Premium users can publish their app as a Progressive Web App (PWA) on a .northerngo.com subdomain from My Projects → More → Create PWA. Pro users can go further and connect their own domain. SSL is issued after the domain is also added under Custom Domains on the Render service.

How does Local AI (WebGPU) work, and are there any limits?

Local AI is available on every plan: you download a model to the browser cache once, and from then on it runs on your computer's GPU. Generation does not use monthly cloud credits, including on Free. The prompt stays on the machine while the local model is writing code. Every model is free on every plan: Ministral 3 Reasoning 3B, Qwen 3.5 4B (the default), Qwen 2.5 Coder 7B, Llama 3.1 8B and Qwen 3.5 9B. The larger ones were Pro until September 2026; the only thing that limits you now is free graphics memory, from roughly 3–4 GB for the 3B model to roughly 7 GB for the 9B. Unlimited means unlimited number of local runs, not unlimited quality. The local models are 3–9 billion parameters; they are considerably weaker than Gemini on large or compound tasks, and on a typical laptop they are often slower than the cloud path. On Premium or Pro, a new local build may still ask the cloud Architect for a structure sketch when you are online — that step is skipped offline, on Free, and on updates. What a Premium or Pro subscription pays for is cloud generations, hosting, subdomains, ZIP export and integrations — not the local models, which are free. Hardware, first download size and the quality gap are covered in the questions above.

How does the AI Agent team work?

The cloud path uses three specialised roles. The Architect plans the structure, the Builder writes the code, and the Critic reviews mechanical damage before you see it. The Critic is not a guarantee that the app does what you meant. The first time you press Generate on a new idea, the Architect asks one follow-up question instead of writing the app. That is deliberate: it pins down who logs in and what must be stored, which is what makes the first build usable. After you answer, the Builder starts. See “What does the Architect do”. Locally the Builder runs on your GPU. A short Architect follow-up question can also run locally. On Premium or Pro, a new local build may still ask the cloud for a structure sketch when you are online. There is no LLM Critic rewrite on the local path — a syntax check blocks JavaScript that will not parse; it does not repair logic.

What does the Architect do, and why is the first reply a question?

The Architect is the first of the three cloud agents. On a brand-new idea it does not write the app. It asks one short question — who logs in, what must be stored, what the home screen must show — so the Builder is not guessing. The question does not use one of your monthly cloud generations. Those are used when the Builder writes the app. Answer in the same box and press Generate again. Later changes to an existing app skip the Architect and go straight to a patch. The first finished app is then saved to your account automatically, so a refresh does not throw the preview away.

How do I connect my local Ollama model?

Pro only. Generation then runs on your machine, with no cloud meter. 1. Install Ollama from their official site and start it with NorthernGo allowed as an origin, so the browser can reach it: Mac/Linux: OLLAMA_ORIGINS="https://northerngo.com" ollama serve Windows (PowerShell): $env:OLLAMA_ORIGINS="https://northerngo.com"; ollama serve 2. In the account menu, open Local AI (Ollama) under Premium / Pro. The dialog tests the connection as it opens. 3. Under Add a model, pick from the shortlist (grouped by 16, 32 or 60 GB of memory) or paste a Hugging Face link. NorthernGo reads the model card and downloads it into Ollama with a progress bar. You do not have to use the terminal. The terminal still works if you prefer — `ollama pull qwen3.5:4b` is the default. 4. Switch on the toggle and pick the model. Save. Only GGUF text models work. Embedding and vision-only repositories are rejected with an explanation. Anything imported from Hugging Face is marked untested in the builder: it will run, but we have not measured whether it builds usable apps. Avoid setting OLLAMA_ORIGINS to "*" — that lets any website you visit send requests to your local Ollama instance. If in-browser WebGPU is also on, WebGPU wins; turn it off in the prompt box to use Ollama.

Can I use Claude, Cursor or ChatGPT together with NorthernGo?

Yes. On Pro, connect Cursor, Claude Desktop or Codex to NorthernGo's official MCP and ask it to build — generate_app runs the same cloud Generate as the editor (no NorthernGo credit, typically 1–3 min). create_app lands HTML you already wrote, without Generate. Neither tool publishes. See “How do I start a NorthernGo build from Cursor, Claude or Codex”. Without MCP: a northerngo.com/?prompt= link still fills the editor; you press Generate. Import HTML and Own API key (Claude / GPT) are Premium and Pro. Import uses no NorthernGo credits; BYOK skips the NorthernGo meter. Connecting a remote MCP as a data source to an app you already built is a separate Pro feature — see “Can I connect an MCP server to my app”. A Gemini- or MCP-created app can itself become a hosted MCP server — see “Can my generated app become an MCP server that others connect to”.

How do I start a NorthernGo build from Cursor, Claude or Codex?

Pro, logged in. This is NorthernGo's official MCP — not the BYOK key dialog, and not Cursor's GitHub/Slack Integrations page. 1. My Projects → More on any app card → MCP. 2. Under “NorthernGo MCP for Cursor, Claude and Codex”, create an MCP key. It is shown once. Copy the JSON snippet (url + Bearer ngmcp_…). Do not paste a Firebase login token. 3. In Cursor: Settings → MCP (Customize → MCPs), or put the JSON in .cursor/mcp.json. Claude Desktop and Codex use the same Streamable HTTP URL if they speak MCP. 4. In the chat, describe the app in plain language, e.g. “Build a booking app for a salon in Lycksele on NorthernGo.” Cursor should call generate_app, then poll generate_status. That does not use a NorthernGo credit. Typical wait 1–3 min. 5. Open the live ?app= URL that comes back. Publish a subdomain yourself in NorthernGo — MCP cannot publish. If Cursor already wrote the HTML, ask it to call create_app instead (no credit). The “Bokio” and “Remote MCP server” fields on the same panel are a data source for the published app, not the Cursor connection. Bokio steps: see “How do I connect Bokio to my app”. A third, optional switch — “This app as an MCP server” — is documented under “Can my generated app become an MCP server that others connect to”.

Does using my own OpenAI or Anthropic key use NorthernGo credits?

No. Own API key is Premium and Pro. When it is on and the key is valid, that cloud generate skips the NorthernGo meter. You pay Anthropic or OpenAI instead. Your API key is not stored in NorthernGo’s database. With BYOK it is used from your browser for that API call and is not sent to NorthernGo’s AI meter. If WebGPU or Ollama is on, those still run first. A failed BYOK request does not fall back to Gemini.

Can Claude or ChatGPT control my NorthernGo account?

They cannot publish your NorthernGo apps. On Pro they can start a cloud Generate via generate_app — same engine as the editor, no NorthernGo credit. See “How do I start a NorthernGo build from Cursor, Claude or Codex”. Connect Cursor, Claude Desktop or Codex to NorthernGo's official MCP (My Projects → MCP) with an ngmcp_ key. Poll generate_status. create_app lands finished HTML without Generate. get_app and save_app iterate HTML. The server will not publish, and will not call tools/call on your remote MCP. Do not paste a Firebase login token — create an ngmcp_ key. Connecting a remote MCP as a data source to an app you already built is a separate Pro feature under the same panel — see “Can I connect an MCP server to my app”. Turning the generated app itself into a hosted MCP server is a third, optional switch — see “Can my generated app become an MCP server that others connect to”.

How do I test my app's performance and SEO?

You can evaluate your app directly within 'My Projects'. Once your app is saved and live on a PWA subdomain or custom domain, click 'Test' next to the app's ID. This runs a live Google PageSpeed audit measuring Performance, Accessibility, Best Practices and SEO. When the test finishes, four coloured dots appear next to the ID — click them to see the exact scores.

Can I take payments in my app?

Yes, you can build full payment flows on the Pro plan. Paste your Stripe secret key into the project menu — it is encrypted at rest, and you can rotate or remove it at any time. The AI can then integrate checkouts, recurring subscriptions or one-time payment products directly into your application code.

Can I connect my Shopify store?

Yes, Pro plan users can integrate e-commerce functionality. By providing your Shopify domain and Storefront token in the menu, the AI can build custom storefronts, fetch live product data and manage shopping carts directly inside your NorthernGo app.

How do I connect Bokio to my app?

Pro, logged in, on a saved project. This is a hosted adapter on the same MCP panel — not a Streamable HTTP URL you paste, and not the ngmcp_ key for Cursor. 1. In Bokio, open Settings → API Tokens (https://app.bokio.se/settings-r/private-integrations) and create a private integration. Copy the token once. 2. The company ID is the UUID in the Bokio URL while you are signed in on that company (you can paste the whole URL). 3. In NorthernGo: My Projects → More on the app card → MCP. Under “Bokio (private integration)” paste company ID and token, then Save Bokio. 4. Rebuild the app (Generate). A Pro rebuild reads the live tool names into plan and logic — bokio_create_invoice (customerRef, lineItems), bokio_create_journal_entry, bokio_list_customers, and the rest. Views do not get that catalog. 5. Open the live ?app= URL or the published subdomain. The editor preview does not reach Bokio. The published app calls window.NorthernGoMCP.callTool — do not write your own fetch to api.bokio.se, and do not put the token in the HTML. Empty both fields and save to disconnect. If a generic remote MCP URL is also saved, the Bokio adapter wins for list and call. Bokio’s own rate limits apply (see docs.bokio.se). This is a private integration, not Bokio’s public OAuth app. Fortnox is still a webhook/Zapier pipe, not this adapter. See also “Can I connect an MCP server to my app”.

How do I connect Gmail to my app?

Pro, logged in, on a saved project. This is a hosted adapter on the same MCP panel — not a Streamable HTTP URL you paste, and not the ngmcp_ key for Cursor. 1. My Projects → More on the app card → MCP. Under Gmail, click Connect Gmail. 2. Sign in with the Google account that owns the mailbox. NorthernGo asks for gmail.readonly and gmail.send. Google may show “Google hasn’t verified this app” — that is Gmail’s restricted scopes, not a NorthernGo failure. Click Advanced, then the grey “Go to northerngo-backend.onrender.com (unsafe)” control. That is the OAuth continue. Do not open northerngo@googlegroups.com — that Google Groups mail opens Outlook or another mail app and never connects Gmail. Success lands on northerngo.com/?gmail=ok (failure: ?gmail=error). The refresh token is encrypted in gmail_connections — it is not written into the HTML. 3. Rebuild the app (Generate) after connecting. A Pro rebuild reads the live tool names into plan and logic — gmail_list_messages (q), gmail_get_message (id, format), gmail_send (to, subject, body). Views do not get that catalog. 4. Open the live ?app= URL or the published subdomain (name.northerngo.com). The editor preview does not reach Gmail. The published app calls window.NorthernGoMCP.callTool — do not write your own fetch to gmail.googleapis.com, do not put googleClientSecret in the HTML, and do not use window.NorthernGo.sendEmail for this mailbox (that path is Resend). Subdomain and custom domain use the same MCP backend as ?app=. Disconnect on the same panel. This is the project owner's mailbox, not each end-user's Gmail. Visitors of the live app read and send as that connected account. A generated app can call window.NorthernGoAI.generate (Premium/Pro in-app AI) to summarise a message or draft a reply, then gmail_send. That sends a new mail — not a Gmail thread reply. Cursor’s ngmcp_ key cannot call gmail_*; those tools run in the published app, not from generate_app. See also “How do I connect Bokio to my app” and “Can I connect an MCP server to my app”.

Can I connect an MCP server to my app?

Yes, on the Pro plan. Three different things share that panel: a remote MCP as a data source for an app you already built; NorthernGo's official MCP so Cursor, Claude Desktop or Codex can generate_app or create_app; and an optional switch so this app itself is a hosted MCP server. The first two do not publish. For the Cursor recipe see “How do I start a NorthernGo build from Cursor, Claude or Codex”. For Bokio (company UUID + private integration token, no MCP URL to paste) see “How do I connect Bokio to my app”. For Gmail (OAuth, project owner's mailbox) see “How do I connect Gmail to my app”. To attach a generic data source: Open My Projects, click More on the project card and choose MCP. Paste the HTTPS address of a remote MCP server that speaks Streamable HTTP, and an optional bearer token, then save. The published app calls window.NorthernGoMCP.listTools() and .callTool(name, args). The editor preview does not reach that server; save the project and open the live ?app= link or subdomain. On a Pro rebuild, the builder reads the live tool list into the plan and logic prompts so the model uses real names. View prompts stay without that catalog. If the MCP is unreachable the build continues without it. Local stdio servers (npx …) cannot run from a browser app. Data sent to the MCP URL leaves the default NorthernGo database path; you are the controller for what that server does with it, the same as with a webhook. The same panel can issue a long-lived ngmcp_ key for Cursor, Claude Desktop or Codex against https://northerngo-backend.onrender.com/api/ng-mcp. That official MCP can generate_app (same cloud Generate as the editor, no NorthernGo credit, poll generate_status) or create_app (land finished HTML without Generate). It cannot publish, or call tools on your remote MCP. Do not paste a Firebase login token there. A third, optional switch on that panel turns this app as an MCP server: other agents connect to /api/app-mcp/{projectId} with a separate ngapp_ key and list/save/delete the app’s NorthernGoDB collections. Off by default. Pro only. Only apps created with Gemini Generate or via MCP create_app / generate_app — not local Ollama or WebGPU. Do not reuse the ngmcp_ key there.

Can my generated app become an MCP server that others connect to?

Yes, optionally, on Pro. It is off by default. After a Gemini Generate in the editor, or after MCP create_app / generate_app, open My Projects → More → MCP and use “This app as an MCP server”. That hosts list/save/delete tools for the app’s NorthernGoDB collections at https://northerngo-backend.onrender.com/api/app-mcp/{projectId}. When Gmail, Bokio or a remote MCP is connected to the same app, those tools are listed there too, with the same names and arguments as window.NorthernGoMCP.callTool. A read-only key only sees read tools. You create a separate ngapp_ key (shown once). Do not reuse the platform ngmcp_ key. Local Ollama and WebGPU apps are not eligible. Writes (save/delete) stay off until you allow them. Anyone with the ngapp_ key reads the rows as the owner — treat it like a service key. The generated HTML is not itself a server; NorthernGo’s backend speaks MCP on the app’s behalf. See also “Can I connect an MCP server to my app” for the other two MCP uses on that panel.

Can I embed the app on my current website?

Yes, Premium and Pro users can generate an embed widget. The platform provides an HTML script tag you can paste into WordPress, Wix, Squarespace or any custom site to render your NorthernGo app inside that page.

How do I manage the user data collected in my app's database?

You control the data settings from the project menu. With a single click you can wipe all collected user entries — emptying the cloud database while keeping the app configuration intact — or reset admin authentication passwords. These tools exist to help you meet your own GDPR obligations as the data controller for your app; the responsibility for what you collect and why remains yours.

How do I change or cancel my subscription?

Click your profile icon and choose 'Manage Subscription'. You are redirected to our Stripe customer portal where you can upgrade, downgrade, update payment details or cancel. Account deletion, including all associated data, can also be run from the same profile menu.

Build it yourself

NorthernGo turns a plain-text description into a working web app with a database, login and a live URL. Local AI generation runs on your own GPU, is unlimited, and is free on every plan.

Start building free