AI

Cobblr uses AI for a few assists, but only if you turn it on with your own provider. Nothing is sent anywhere until you connect one.

1. Open Configuration → AI (or your own account, for a personal connection) and press Connect. The button says a free tier is available.
2. Pick a provider. The free option leads, with numbered steps and a link that opens the right page.
3. Paste the key. It is checked as you paste, and you get Works or the reason underneath.

What AI helps with

• Identifying what you scan. For a photo with no barcode, AI reads the image. For a barcode, Cobblr looks the code up in public product databases first, and pulls in AI to name it when the databases don't recognize the code or come back with only a bare listing. Either way you get a name to confirm.
• Drafting a bundle from a plain-language description of what you want to track.
• Chatting with Cobb, the assistant. Cobb reads your records to answer, and proposes changes (create, update, delete, run an action) that you confirm before anything is written.
• Smaller assists around the app.

With no AI connected

These features work in a simpler mode rather than disappearing, and the rest of Cobblr works exactly the same:

• Scanning still captures items into the inbox and still resolves barcodes from the product databases. What you lose is the smart part: reading a photo that has no barcode, and fleshing out a bare listing.
• Bundle building still works from a description. Cobblr compiles a prompt you copy into your own ChatGPT or Claude, then paste the result back for the kernel to validate. It just isn't automatic the way a connected provider makes it. Starting from the ready-made catalog is always there too.
• Cobb the assistant drops to basic mode, which is more than canned phrases: it answers counts and "where is my drill" from your own records, clears duplicates and fixes broken links, and carries out whole instructions the workspace has been taught. What it cannot do is understand a sentence nobody has taught it.

Nothing requires AI, and the page says so plainly rather than making you find out. But scanning and Cobb are both noticeably better with it, so connecting a provider is worth the five minutes. The setup leads with the free option for that reason, and where the page tells you nothing is connected it gives you a Connect button rather than leaving you to go looking.

Adding a provider

Picking a provider shows what it costs, numbered steps to follow, and a link that opens the right page. Google AI Studio, Ollama and Anthropic each carry their own:

• The Google steps end with pasting the key.
• The Ollama steps cover installing it and pulling a model that can read photos.
• The Anthropic steps start with adding credit, because a new account has none and a key without credit fails on first use.
• Caveats are shown up front rather than buried in a step: that Google's free tier uses what you send to improve their products, that a local model is less accurate but never sends your photos anywhere.

Choosing a provider is the fuller comparison, including the one setting to get right on the free Google tier.

The key is checked as you paste it

• Works, or the reason it did not, appears under the field about half a second after you stop typing. A wrong key usually reads status 400.
• Nothing wrong is stored. The check runs on what is currently typed, before anything is saved, so a bad key never becomes a broken connection you have to go back and fix.
• The Model field becomes a dropdown of the models your key can actually reach. Leave it on Use the default unless you want a specific one. The default is named on the field so you can see what you are getting.
• Chat models are grouped first. The dropdown puts them under Chat models and the rest under Everything else this key can reach. Nothing is hidden.
• All three places you can add a key do this: adding a provider to a workspace, replacing its credentials, and your own connections under Your account.

Why the list has two groups, and what a failed check can mean

Providers do not curate that list. A Google AI Studio key reports 51 models, and most of them generate video, music, speech or embeddings and cannot answer a prompt at all. The second group is not junk, it is the models Cobblr could not identify from the name, and it is there precisely so a model it does not recognise is still selectable.

If the check cannot reach the provider at all, read that as a network answer rather than a key answer: a self-hosted model on your own machine has to be reachable from Cobblr, which for a hosted workspace means through your edge bridge.

The same steps appear whether you are setting AI up for a workspace under Configuration or just for yourself under your account. Your browser will not autofill a saved password into the connection fields, since a key is not a password and a silently filled one fails in a way that is hard to read.

Several personal connections on one workspace

You can route more than one of your own connections to a workspace, and then choose which goes first.

• The workspace row says what already powers it ("Ollama is already powering this workspace") and offers Use this one first or Keep the other first.
• Nothing falls through to a second connection. When the first errors or hits its daily cap, the call fails and says so rather than silently spending a different key.
• The order is per kind of work, so a fast model can serve the live camera scan while a slower, better one handles inbox identification and chat.

Why you choose, instead of the newest edit winning

Before, the connection you had edited most recently quietly won, so renaming or re-keying one could change which model answered in a workspace you had not touched.

Bring your own provider

AI runs on a provider you connect, using your own key or your own server:

• A hosted provider like OpenAI, Anthropic, or OpenRouter, with your API key.
• A local model you already run, by pointing Cobblr at your own Ollama or any OpenAI-compatible v1 endpoint (LM Studio, vLLM, llama.cpp, and the like). This is how you use whatever model you run somewhere else, and nothing leaves your network.

Cobblr ships no built-in AI service, so there's no shared quota and no cost you didn't sign up for. Connect one (for a workspace, or as a personal connection you grant to the workspaces you choose) and the AI features light up.

Turning AI off for one workspace

Owners and admins can switch AI off per workspace. On Configuration → AI, turn off Use AI in this workspace.

• That workspace runs in basic mode: Cobblr makes no AI calls on its behalf, even when a provider is available on the instance.
• A member who connects their own personal key can still use it there.

What the status banner says

The status banner on that page is honest about where AI stands. When AI is on, it says where it comes from: a key the workspace added, or a personal connection shared into it. When you have turned it off but a provider is still connected, the page says AI is there to use and that flipping the switch back on will use it, so "off" reads as a choice rather than a dead end.

Reviewing AI activity

Every AI call is logged under Configuration → AI.

• Monthly totals and per-capability rows for calls, spend, and failures.
• Open a call to see the prompt and the response as labelled, readable sections, with chat turns broken out and any image the call carried shown. The raw JSON stays one click away.
• "tokens not reported" is what the log says when a provider does not report token counts, instead of showing a zero it never claimed.

Turning it off entirely

On a self-hosted install, an operator can disable all built-in AI with one switch, COBBLR_AI_ENABLED=false, no matter what keys are set. That's the hard floor for an instance that should never make an AI call.

Privacy

When AI is on, the content it needs (a photo, a description) goes to the provider you chose, not to Cobblr. With no provider connected, nothing leaves the box. See Self-hosting -> Privacy.