Try the AI for free (OpenRouter)

Free models cost nothing per token. What you give up is throughput and, usually, some capability:

• Rate limits — free models are capped per minute and per day against your OpenRouter account. The numbers are OpenRouter's, they change, and orgadmin.ai neither enforces nor tracks them, so check the current limits in your OpenRouter account rather than trusting a figure you read somewhere.
• What hitting a limit looks like here — a run fails outright, or an answer stops halfway through. The app does not queue or retry for you. Wait a moment and run it again.
• Smaller models, shorter context — the free tier skews towards smaller or older models, and several have context windows too short to hold a large org's schema.
• The roster moves — models get added, renamed and withdrawn regularly. A model id that worked last month may simply not be in the list today.

How the key itself is stored, and what changes once you move to a paid model, is in Bring your own AI key.

1. 1

   Create an OpenRouter key

   Sign up at openrouter.ai and create a key under Keys. It is your account with them — orgadmin.ai never creates one for you.
2. 2

   Open the AI card

   Go to Settings and find Your AI provider. If you have never added a key it reads No AI key added yet — click Add your API key. If you already have one, use the pencil Edit key icon.
3. 3

   Pick OpenRouter and paste the key

   Set Provider to OpenRouter — the muted line underneath should change to Key from: openrouter.ai → Keys. Paste the key into API key, and leave Base URL empty; OpenRouter does not need one.
4. 4

   Fetch the model list

   Click Fetch models. This step matters, because the model picker is empty until you do — until then it just says No models loaded. A toast tells you how many were found; for OpenRouter it is several hundred.
5. 5

   Find a free one

   Open Model and type in the search box. Searching :free lists the whole free set; searching nemotron jumps to NVIDIA's family. nvidia/nemotron-3-super-120b-a12b:free is a good first pick — free, a large context window, and it handles tool calling, which the app depends on. If the roster has moved on and it is not there, any other :free id is worth a try. Already know the id? Click Enter ID manually and paste it.
6. 6

   Save

   Click Save key. Saving runs a quick tool-calling check — the next section covers what to do if it complains.

When you save, orgadmin.ai sends the model one trivial request that can only be answered by calling a tool. Weak models often fail it — they reply in prose instead, or print the tool call out as literal text like <function=…>. This is the most common problem with free models, and it is worth understanding because the failure is a quiet one.

Tool calling is not optional here. Chat, the SOQL assistant, Insights and Studio all work by calling tools to read your org — objects, fields, Apex, flows, query results. A model that cannot call them does not fail loudly; it answers from thin air, confidently and wrongly.

• “This model may not support tool calling” — it answered without using the tool, or wrote the call out as text. Click Edit key, pick a different :free id and save again. Free models vary enormously, so try two or three before concluding it cannot be done.
• “Couldn't verify tool calling” — a different outcome: the check timed out or the provider returned an error, which on a free model is often just a rate limit. Save again in a minute and it may pass.

If nothing in the free list passes, that is the honest signal to spend a little on a small paid model rather than keep fighting it.

This is the part to understand before you judge the app on a free model. orgadmin.ai does the plumbing: it pulls your metadata, keeps it fresh, assembles the right context, defines the tools, and renders the result. Every act of judgement in between — reading the schema, deciding what to look up next, spotting the actual problem, writing the summary — belongs to the model, not to us.

So the quality of every AI answer you get is bounded by the model you picked. Nothing in the app makes a weak model smart. When answers come back shallow, plainly wrong, or trail off before finishing the job, that is almost always the model rather than the feature.

Features do not lean on the model equally. Roughly hardest to easiest:

• Toughest — Security Review and Tech Debt Finder: long multi-step reasoning across a lot of metadata. Tech Debt makes many calls in a single run, so free-tier rate limits bite here first, and weaker models tend to stall or return generic filler. Flow Health's AI summary uses the same model as Tech Debt.
• Demanding — Development Studio and Insights: writing Apex or LWC that actually compiles, and composing a dashboard that returns the right numbers, need both precision and a long context window.
• Most forgiving — AI chat and the SOQL assistant: short, tool-driven turns with a clear question and a checkable answer. Start your test drive here.

• Swapping to a paid model on the same provider is easy. Click Edit key, change the model, and leave API key blank — your stored key is kept. Only changing the provider makes you paste a key again.
• Set a spending limit first. On a paid model every request is billed to your own account, including failed and cancelled runs. See Bring your own AI key.
• Keep an eye on consumption. The Usage page tracks what the app spent on your key — see Usage & billing.
• Then go and use it. AI chat is the best first stop, and Your first 15 minutes is the guided tour if you have not done it yet.