Calik AI does not resell tokens. You connect a model provider directly with a key that
belongs to you, and the whole workspace runs on it: 94 models, 48 agents, 43 tools.
This page is the setup guide for all 27 providers the app accepts, starting with
Claude, and five of them hand out working keys without asking for a card.
Whichever you pick, the key is stored encrypted on your device and your prompts go
straight from you to the provider. We never see either one.
How to get a Claude API key
Anthropic is the provider most people arrive here looking for, so it gets the full walk
rather than a row in a table. Ten minutes end to end, and about five of those are the
phone verification.
Anthropic Console Prepaid
Claude Opus 5, Claude Fable 5, Claude Sonnet 5, Claude Haiku 4.5
Open platform.claude.com
and sign up with an email address or a Google account. This is the developer console,
which is a different product from the Claude chat app. The old address, console.anthropic.com, now redirects here.
Verify your phone number when it asks. Anthropic uses it to stop one person opening
many accounts, and you cannot create a key until it is done.
Go to Plans & Billing and add a payment method, then buy credit. The Claude
API has no free tier: it runs on prepaid balance, and the smallest top up is a few
dollars. That balance is the only thing you ever pay for, and you pay it to Anthropic,
not to us.
While you are in billing, set a monthly spend limit. It is the one setting that
turns an API key from an open tap into a budget, and it takes one click.
Open Settings → API keys
and press Create Key. Give it a name you will recognise later, such as
calik-ai, and leave the workspace on Default.
Copy the key the moment it appears. It starts with sk-ant- and Anthropic
shows the full value exactly once. If you lose it, you delete that key and make
another one, which costs nothing.
In Calik AI, open API keys, choose Anthropic, paste and save. The model
picker fills with Claude Opus 5, Fable 5, Sonnet 5 and Haiku 4.5 straight away.
A Claude Pro or Max subscription is not an API key. The chat plan on claude.ai and
the API are billed separately, and neither includes the other. If you already pay for
Pro you still need console credit to use Claude here, and the console will not accept
your claude.ai session token as a key.
Which Claude model to start on
Haiku 4.5 for volume work, classification and quick drafts: it is the cheapest per
token by a wide margin. Sonnet 5 is the sensible default for everyday work.
Opus 5 for the problems where a wrong answer costs more than the tokens do, which
in practice means long refactors, architecture calls and anything you would otherwise
have to check by hand. Switch per conversation in the app, so the expensive model is
only in the room when it needs to be.
Rate limits, and why they grow
Anthropic puts new accounts on a low usage tier and raises it automatically as your
cumulative spend passes each threshold. A fresh key will hit 429 sooner than
you expect on a busy Agent Mesh run. That is not a fault in the key and there is nothing
to fix: either pace the run, or let the tier climb on its own.
Keeping the key safe
Treat it like a password. It goes into your device's encrypted storage here, never into
our database or our backups. Do not paste it into a chat message, a screenshot or a
public repository. If it ever leaks, revoke it in the console and create a new one:
revocation is instant and the old key stops working the same second.
Five routes that cost nothing
Every one of these ends with a key the app already knows how to use, and none of them asks for a card.
01
Google AI Studio
Fastest
Gemini 2.5 Pro, 2.5 Flash and 2.0 Flash. A Google account is the only requirement,
and the free tier does not ask for a payment method at all.
Press Create API key. If it asks for a project, let it make a new one.
Copy the key. It starts with AIza.
In the app, open API keys, choose Google AI (Gemini), paste, save.
The catch: requests are rate limited per minute and per day, and Google's free
tier terms allow them to use the content for improving their products. Fine for
drafting, comparing and learning. Not the route for anything confidential, which is
what route 05 is for.
02
Groq
Quickest answers
Llama 3.3 70B, Mixtral 8x7B and Gemma2 9B, served on custom silicon that returns
tokens faster than anything else on this list. Free tier, no card.
Copy it while it is on screen. It starts with gsk_ and is shown once.
In the app, open API keys, choose Groq, paste, save.
The catch: open-weight models only, so no GPT or Claude here, and the free tier
has per-minute request limits. The speed makes it the best pick for Agent Mesh, where
several models answer at once.
03
GitHub Models
If you code
GPT-4o, Llama 3.1 405B, Mistral Large and Phi-3.5, through a normal GitHub personal
access token. If you already have a GitHub account you are most of the way there.
In the app, open API keys, choose GitHub Models, paste, save.
The catch: the free allowance is meant for evaluation, so the request limits are
the tightest here. It is the only free route on this page that reaches a GPT-class
model, which is what makes it worth the ceiling.
04
OpenRouter
Widest choice
One key, a hundred models behind it, and a set of them priced at zero. A single
account gets you several free open-weight models including reasoning ones.
In the app, open API keys, choose OpenRouter, paste, save.
When picking a model, choose one whose name ends in :free.
The catch: only the models marked free cost nothing, and which ones those are
changes over time. Anything else draws on a balance you would have to top up, so
check the label before a long session.
05
Ollama, on your own machine
Nothing leaves
Llama 3, Mistral, Phi-3 and Qwen running locally. No key, no account, no request
limit, and no third party ever sees the text. Free in the strongest sense of the word.
In a terminal, run ollama pull llama3 and wait for the download.
Leave it running. It listens on http://localhost:11434.
In the app, open API keys, choose Ollama (Local) and confirm the address.
The catch: your own hardware sets the ceiling. A laptop with 16GB will run an 8B
model comfortably and will struggle above that. Slower than the hosted routes, and the
only one where nothing you type ever crosses the network.
All 27 providers the app accepts
Read straight out of the app's own provider list, so every console link below leads to the
key format the app expects. Anything without a Free mark needs a
payment method or prepaid credit at the provider, never at us.
Frontier modelsThe labs that train their own frontier models. You buy tokens from them directly.10
Fast inferenceOther people's open-weight models, served on hardware built for speed. Where Agent Mesh pays off, because several models answer at once.06
The same four steps for all 27. Nothing is configured anywhere else.
Open the workspace and find API keys in the settings panel.
Pick your provider from the list. The hint under the field shows the prefix that provider's keys start with, so you can tell at a glance whether you pasted the right one.
Paste and save. The app writes it to your device's encrypted storage, checks it against the provider once, and marks the provider as connected.
Open the model picker. Only the models your connected providers actually serve are listed, so there is nothing in there you cannot run.
Add as many as you like. The app keeps one key per provider and they do not
interfere with each other. Agent Mesh then puts models from different providers in the
same room on the same question, which is the point of holding more than one key.
Which route should you take
Short version: Claude if you want the strongest reasoning and will pay for it, Google AI Studio to start for nothing, Ollama if the text is sensitive.
Route
Models you get
Needs
Best for
Free
Anthropic
Claude Opus 5, Fable 5, Sonnet 5, Haiku 4.5
Card, prepaid credit
The hardest reasoning and long code work
No
Google AI Studio
Gemini 2.5 Pro, 2.5 Flash, 2.0 Flash
Google account
Starting in under a minute
Yes
Groq
Llama 3.3 70B, Mixtral, Gemma2
Google or GitHub account
Agent Mesh, where speed compounds
Yes
GitHub Models
GPT-4o, Llama 3.1 405B, Mistral Large
GitHub account
Reaching a GPT-class model at no cost
Yes
OpenRouter
Several open models marked free
Any account
Trying many models behind one key
Partly
Ollama
Llama 3, Mistral, Phi-3, Qwen
Your own machine
Confidential work, unlimited use
Yes
Nothing stops you using several at once. The app keeps a key per provider, so a common
setup is Claude for the hard questions, Gemini for long context and Llama on Groq for speed.
Agent Mesh puts all three in the same room to argue a question out: three model families,
one answer that has been checked by each of them.
When a key does not work
Nearly every failure is one of these seven, and none of them is a fault in the app.
What you see
What it means
What to do
401invalid api key
The key is mistyped, revoked, or belongs to a different provider than the slot you pasted it into.
Copy it again from the console. Watch for a trailing space, and check the prefix matches the hint under the field.
credit balance is too low
A prepaid provider with an empty balance. Anthropic and DeepSeek both work this way.
Top up in that provider's billing page. Nothing on our side is involved.
429rate limit
Too many requests or tokens per minute for your usage tier. Common on a fresh key during a busy multi-agent run.
Wait a moment and retry, run fewer agents at once, or let the tier rise as your spend does.
insufficient quota
OpenAI's wording for billing that was never set up, or a trial grant that expired. It is not a rate limit.
Add a payment method in the OpenAI billing page and buy credit.
403model not found
Your account cannot see that model yet. Some providers gate their newest or largest models behind a spend threshold or a separate opt in.
Pick another model from the same provider, or request access in the console.
529 or 503overloaded
The provider is saturated. Nothing to do with your key.
Retry shortly, or send the question to a second provider you have connected.
connection refused on localhost
A local route where the server is not running, or is on a different port than the app expects.
Start Ollama, LM Studio or llama-server, then confirm the address in the key field matches the port it printed.
What free actually means here
Free tiers are real, and they are not charity. Providers give them away to get developers
onto their platform, and they pay for that in two currencies: limits and data.
Limits. Every hosted free route above caps how many requests you may send per minute
and per day. For one person working normally that ceiling is rarely reached. For a long
unattended job it will be. The numbers move without notice, which is why they are not
printed on this page: each route links to the provider's own page, and that page is
always right.
Data. Several free tiers reserve the right to use what you send for improving the
provider's models. That is a fair trade for a free service and a bad trade for a client
contract. If the text must not be used that way, either move to that provider's paid tier,
where the terms are usually different, or run the model on your own machine.
Whichever you choose, the arrangement on our side does not change: your key is stored
encrypted on your device, your prompts go straight from you to the provider you picked,
and we never see either. That is set out in full in the privacy policy.
Why we can tell you this and most cannot
Most AI workspaces buy tokens wholesale and sell them back to you with a margin. Telling
you how to hold your own key would be telling you how to stop paying them, so they do not.
We are built the other way around. Calik AI is the workspace, not the model: 27 providers,
94 models, 48 agents and 43 focused tools, all of them driven by a key that belongs to
you. We never touch the inference bill, so a visitor who runs entirely on free tiers costs
us nothing and is welcome to stay there indefinitely, and a visitor who spends a fortune at
Anthropic spends every cent of it with Anthropic. The
paid plan exists for people who want the whole toolkit and the
multi-agent surfaces, not for people who want a token allowance.
That is also why the free tier here has no expiry. There is no point putting a clock on
something that costs us nothing to run.
Questions people ask first
Is the Claude API free?
No. Anthropic runs the API on prepaid credit, so you add a payment method and buy a balance before the first request. There is no free tier and no trial clock. If you want Claude-class quality at no cost, the closest free routes on this page are GitHub Models, which reaches GPT-4o, and Google AI Studio, which reaches Gemini 2.5 Pro.
I already pay for Claude Pro. Can I use that here?
No, and this catches most people out. Claude Pro and Max are subscriptions to the claude.ai chat app; the API is billed separately on prepaid credit. Neither includes the other, and a claude.ai session token is not an API key. You need a key created in the Anthropic Console, which starts with sk-ant-.
How much will a Claude key actually cost me?
It depends entirely on which model and how much text, which is why no figure is printed here. Anthropic's pricing page carries the current per-token rates. The practical control is in your hands: set a monthly spend limit in the console, keep Haiku 4.5 for volume work and bring Opus 5 in only for the questions where being wrong costs more than the tokens do.
Do I have to give a credit card anywhere?
Not to us, ever. Whether a provider asks depends on the provider: Google AI Studio, Groq, GitHub Models and Ollama do not, OpenRouter only asks if you pick a model that is not marked free, and the frontier labs including Anthropic and OpenAI all do.
Where is my API key stored?
In your device's encrypted storage, and nowhere else. It is not copied to our servers, it is not in our database and it is not in our backups. Prompts travel from your device to the provider you chose without passing through us.
Can I connect several providers at the same time?
Yes, and it is the setup we would recommend. Each provider has its own key slot and they do not interfere. Agent Mesh then lets models from different providers answer the same question, critique each other and converge on one result, which is worth more than any single model on its own.
Which free route gives the strongest model?
GitHub Models, because it reaches GPT-4o, though it also has the tightest request limits. For day to day work Gemini 2.5 Flash through Google AI Studio is the better balance of quality and headroom.
What if I do not want anything leaving my computer?
Use one of the three local routes: Ollama, LM Studio or llama.cpp. The model runs on your own hardware, so no prompt, file or answer crosses the network at all. They are slower than the hosted routes and limited by your machine, and they are the only arrangement where privacy does not depend on trusting anyone's terms.
My key stopped working. What broke?
Almost certainly the provider, not the app. An empty prepaid balance, a rate limit on a fresh account, a revoked key or a model your account cannot see yet cover nearly every case. The table above matches the exact wording of each error to the fix.
Then what does the paid plan buy?
The full workspace rather than more tokens: every one of the 43 AI Center modules, the larger Agent Office and Agent Mesh configurations, and the tools kept out of the free tier. The models stay yours either way. See pricing for the exact split.
Get a key, then open the workspace
The free routes take about a minute. Claude takes ten. Either way, when the key is in,
everything else on this site is already unlocked.