BYOK AI explained: bring your own AI key, the why and how.

Updated 2026-04-27 · 11 min read

I've spent a lot of time the last year explaining BYOK to founders who were stacking $20/month subscriptions and wondering why their tooling bill kept growing. The short answer is that they were paying a margin on AI inference twice: once to the tool that bundled it, and once more in the form of a vendor lock-in tax that compounded as their data piled up.

BYOK fixes both of those, with one real trade-off (you have to set up an API key once). This page walks through what BYOK actually means, what it costs in practice, why it changes the privacy math, and how to set up your first key in about five minutes.

What BYOK actually is

Most AI tools come in one of two shapes:

• Bundled inference. You pay one flat fee, the tool handles AI calls behind the scenes, and you don't see the underlying API. Examples: ChatGPT Plus ($20/mo), Claude Pro ($20/mo), Notion AI ($20/mo on top of your Notion subscription), Cursor Pro ($20/mo), Reflect ($15/mo).
• BYOK. You sign up for an account at the AI provider's developer portal, generate an API key, paste it into the tool. The tool makes API calls using your key, and the AI provider bills you directly for usage. The tool charges you separately, typically a one-time fee or a low subscription. Examples: Projelli ($49 one-time + your API costs), Cursor (BYOK option), Continue.dev (free + your API costs), Raycast AI, several Obsidian AI plugins.

The acronym BYOK shows up in the cloud-software world ("bring your own key") to mean a customer-managed encryption key. In the AI tools context, it means the same idea applied to your AI provider account: you're the customer of the AI provider, not of the tool wrapping it.

It matters because it inverts who owns the relationship. With bundled inference, you're a customer of Notion / OpenAI / Anthropic-via-the-tool. With BYOK, you're a customer of Anthropic / OpenAI / Google directly, and the tool is just an interface.

The cost math

Here's what BYOK actually costs in practice, from real usage logs.

Light user: 30 minutes/day of AI conversations, ~5 messages

Typical token usage: ~80,000 tokens/day in, ~40,000 out. With Claude Sonnet 4.5 ($3 per million input, $15 per million output), that's:

• Input: 80K × $3/1M = $0.24/day
• Output: 40K × $15/1M = $0.60/day
• Daily total: ~$0.84
• Monthly: ~$25 if you work every day, ~$17 if you work weekdays only

With OpenAI's GPT-5 mini or Google's Gemini 2.5 Flash, both significantly cheaper for similar quality on most tasks, the same volume runs $5-12/month.

Heavy user: 2 hours/day of AI work, longer context

~250,000 tokens/day in, ~120,000 out. Same Claude Sonnet 4.5 pricing:

• Input: 250K × $3/1M = $0.75/day
• Output: 120K × $15/1M = $1.80/day
• Daily total: ~$2.55
• Monthly: ~$50-75

Heavy users on the cheaper Gemini Flash or GPT-mini tier: $15-30/month.

Comparison to flat-rate alternatives

Option	Light user / mo	Heavy user / mo	3-year total (light)
ChatGPT Plus	$20	$20 (with throttling)	$720
Claude Pro	$20	$20 (with throttling)	$720
Notion + Notion AI	$10 + $10 = $20	$10 + $10 = $20	$720
Cursor Pro	$20	$20 (with throttling)	$720
Projelli (Pro $49) + BYOK Claude	$49 + $5-15 = year 1: $109; ongoing: $5-15/mo	$49 + $30-50 = year 1: $409; ongoing: $30-50/mo	~$229 (light user)
Projelli (Lifetime $99) + BYOK Gemini	$99 + $3-8 = year 1: $135; ongoing: $3-8/mo	$99 + $15-30 = year 1: $279; ongoing: $15-30/mo	~$252 (light user)

The flat-rate tools win in one specific case: you genuinely use AI for many hours a day, every day, without throttling. For most indie founders, BYOK comes out cheaper across any reasonable usage pattern.

Important caveat: usage varies. The numbers above are realistic for a founder who does most of their planning, drafting, and review with AI assistance. Lower-touch users can run as low as $1-3/month. Higher-touch users (running long workflows daily, processing customer interview transcripts) can hit $50+/month.

The privacy story

The cost angle is what gets attention. The privacy angle is more important and less talked about.

Three things change when you switch from a bundled-inference consumer tool to BYOK on the same AI provider's API tier.

1. The AI provider treats your data differently

Consumer tier (ChatGPT.com, Claude.ai, Gemini app) and API tier (BYOK) have different default policies, even though the same company runs both. In short:

• Anthropic API: does not train on inputs by default (per their commercial terms). Consumer Claude.ai may use opted-in data.
• OpenAI API: does not train on API inputs by default (per their API data usage policy). Consumer ChatGPT does train on opted-in conversations.
• Google AI Studio API: may use prompts to improve products on the free tier; opt-out and paid tier are different. Read the Gemini API terms before assuming a default.

The takeaway: API tier is generally a stricter privacy posture than the consumer chatbot. Confirm the current terms before relying on this; they change.

2. The desktop app never sees your prompts

With a bundled-inference tool, your prompt goes: your machine → tool's servers → AI provider → tool's servers → your machine. The tool's company sees your prompt and the response.

With a BYOK desktop app like Projelli, your prompt goes: your machine → AI provider → your machine. The desktop app's servers are not in the path. There's no intermediate store. Cursor, Continue.dev, Projelli, and most BYOK tools work this way. (Always check; some tools route through their own servers even with BYOK, which is a different threat model.)

3. Your conversation history is on your machine

Even if the AI provider has a strict no-training policy, the question of where your conversation lives still matters. With a bundled tool, the conversation is in their database. With a BYOK desktop app, the conversation is a file on your hard drive, in a folder you picked.

For an indie founder whose AI conversations contain pitch decks, customer interviews, and pricing strategy work, this is the difference between "in someone else's database with a TOS that can change" and "on my machine, period". Most founders prefer the second posture once they think about it.

How to actually set up an API key

Five minutes, three providers. Pick one to start; you can add the others later.

Anthropic Claude

1. Go to console.anthropic.com and sign up
2. Add a payment method and put $5-20 of prepaid balance (you can set a hard monthly cap so you don't accidentally spend more)
3. Go to Settings → API Keys → Create Key
4. Copy the key (starts with sk-ant-)
5. Paste it into your BYOK app (in Projelli: Settings → AI → API Keys → Anthropic)

You're done. Claude conversations now go directly from your machine to Anthropic.

OpenAI

1. Go to platform.openai.com and sign up
2. Add a payment method, put credit on file, and set a monthly usage cap in Settings → Limits
3. Go to API Keys → Create new secret key
4. Copy the key (starts with sk-)
5. Paste into your app's OpenAI settings

Google Gemini (via AI Studio)

1. Go to aistudio.google.com
2. Click "Get API key" in the left nav
3. Create or pick a Google Cloud project (the prompt is straightforward)
4. Copy the key (starts with AIza)
5. Paste into your app's Gemini settings

Gemini has a generous free tier; you may not need to enter a payment method to start. Read the data terms before relying on the free tier for sensitive content.

Setting a hard monthly cap (do this immediately)

All three providers let you set a monthly spend cap. Use it. A typical sequence the first time:

1. Set the cap at $10/month while you learn your usage
2. Use the tool for a week, see what your actual daily cost is
3. Raise the cap to 2-3x your observed monthly usage as a safety margin

The cap is a hard ceiling. Once you hit it, the API stops responding until next billing cycle. This protects you from runaway costs if a tool starts looping (rare but possible).

What about local AI?

You can run Llama, Mistral, or other open-weight models locally with tools like Ollama or LM Studio. This is fully off-cloud: no API key, no provider, no internet required. Privacy maxes out, cost is zero (except electricity), and the data stays entirely on your machine.

The trade-off is quality. Open-weight models are catching up but are not at parity with Claude Sonnet, GPT-5, or Gemini 2.5 Pro for the kind of strategic-thinking work most indie founders want AI for. Drafting code? Open-weight models are competitive. Writing a pitch deck? You'll feel the difference.

The realistic 2026 answer: use BYOK with Claude/OpenAI/Gemini for high-stakes work, and use local Ollama for the volume work where 90% quality is fine. Projelli supports both: configure local Ollama as a provider alongside Claude / OpenAI / Gemini, and route per-conversation.

Common concerns answered

"The 5-minute setup is going to be the part everyone gives up on"

Some people will. They're not the target market for BYOK tools. The audience that buys local-first BYOK apps is exactly the audience comfortable getting an API key. If "5 minutes setting up an account at console.anthropic.com" is a dealbreaker, the bundled-inference tools are a fine choice.

"What if I don't want to attach a credit card?"

Google's Gemini has a generous free tier that doesn't always require a card. Anthropic and OpenAI both require payment info to use the API at all, even with a $5 prepaid balance. If avoiding cards is your priority, Gemini is the right starting point. After a month of using it for free, you'll have a sense of what your actual usage is, and adding a card to one of the others becomes a smaller mental ask.

"What if the API key gets stolen?"

Three layers of protection. First, the desktop app stores the key in your OS keychain (macOS) or an encrypted file (Windows / Linux), not in plaintext anywhere. Second, all three providers let you revoke a key with one click and generate a new one. Third, the monthly spend cap means even a stolen key has bounded damage. Treat the key like a password: don't share it, rotate if you suspect exposure, and you're fine.

"What if the AI provider goes down?"

Same risk as a bundled-inference tool, since they use the same providers under the hood. A BYOK app actually gives you a small advantage: you can quickly switch providers (Claude → OpenAI → Gemini) if one is having an outage. Bundled tools are stuck with whatever they wrap.

Tools that support BYOK today

Tool	BYOK support	Workspace shape
Projelli	Yes, all 3 providers	Local-first AI workspace for founders
Cursor	Yes (Pro tier)	Code editor with AI
Continue.dev	Yes	VS Code AI extension
Raycast AI	Yes (Advanced)	Mac launcher with AI
Obsidian + Smart Connections / Copilot	Yes	Local notes with AI plugins
ChatGPT, Claude.ai, Gemini app	No	Cloud chat, bundled inference
Notion AI	No	Cloud workspace, bundled inference
Reflect, Mem.ai, Tana	No	Cloud notes, bundled inference

This list moves. Some tools are adding BYOK as a Pro option; some are removing it. Check the current state of any tool you're evaluating.

Frequently asked questions