Bring Your Own LLM: Why GitChat Lets You Use the AI Model You Already Trust

If you have used more than one AI developer tool in the last twelve months, you have already noticed the pattern.

Each tool comes with its own opinion about which AI model is best. Each tool ships with a default. Each tool, in subtle ways, ties you to that default. The model gets updated on the vendor's schedule. The pricing reflects the vendor's relationship with the model provider. The capabilities of the tool are bounded by what the chosen model can do.

For some tools this is fine. The integration is tight, the model fits the task, and the user does not particularly care which model is doing the work as long as it works. For others, the model lock-in starts to chafe. The default model is not the one you trust most. The pricing is opaque. The capabilities feel limited because the chosen model is limited.

The teams that have been using AI tools for a while have started to think about this more carefully. They want to know which model is doing what. They want the option to switch. They want their best subscription, the one they are already paying for, to be the one their tools use.

This is the story of why GitChat ships with no default model preference, why we let you bring your own API key, and why we think this is the right architecture for the next decade of AI developer tools.

Why the lock-in happens

Most AI developer tool vendors integrate with a single model provider because it is the path of least resistance. Pick a partner. Ship a product. Optimize for that partner's strengths. The product launches faster. The marketing is cleaner. The integration is tighter.

This is rational for the vendor. It is increasingly unsatisfying for the customer.

The customer wants to use the model that is best for their use case, which sometimes is the vendor's chosen partner and sometimes is not. The customer wants to use the model they have already paid for in another context, which is rarely the same one. The customer wants the option to switch when the model landscape changes, which it does, frequently.

The lock-in is not always intentional. Some tools simply have not built the abstraction layer that would let them swap providers. Others have, but have not made it visible to the user. A few have, but charge extra for the privilege of bringing your own key. The result is the same. Most users end up paying twice: once for the tool, once for whichever model the tool insists they use.

What "bring your own model" actually means in GitChat

GitChat works with any of nine major AI providers. You configure one or more API keys in your account. You choose, on a per-conversation basis, which model handles the work. The supported providers are Anthropic for Claude, OpenAI for GPT-4o and its successors, Google for Gemini, xAI for Grok, Mistral for the Mistral family, OpenRouter for unified access to dozens of models from multiple providers, NVIDIA for hosted open models, LM Studio for fully local models running on your own hardware, and a fallback for any OpenAI-compatible endpoint.

This is not a comparison chart in our marketing. It is the actual list of providers wired into the product. If you pay for Anthropic, your work uses Claude. If you have an OpenAI subscription, your work uses GPT. If your security team requires that no data leaves your machine, your work uses LM Studio against a local model. The same product, the same features, the same workflow. Different model, depending on what you already use.

The implications are significant.

You are not locked into one vendor's pricing. If a competitor releases a better model for less, you switch the key. If your provider has an outage, you fall back. If your team has different preferences for different kinds of work, you use different models for different conversations.

Your data goes to your provider, not to us. We do not see your prompts. We do not see the responses. The AI calls happen through the credential you supplied, directly to the provider you chose, with the relationship you already have with them. The data flow is the same as if you were using the provider's API directly, with the workflow benefits of GitChat layered on top.

Why this matters more than it sounds

The bring-your-own-model architecture is not just a feature. It is a strategic commitment.

The AI model landscape has been changing rapidly for the last three years. Models that were state of the art at the beginning of 2024 are middle of the pack by the end of 2025. New entrants emerge. Pricing shifts. Capabilities expand. The provider that is best today is not guaranteed to be best a year from now.

A tool that locks you into one provider is making a bet on your behalf. The bet is that today's choice will still be the right choice when you decide to renew. Historically, this bet has been wrong as often as it has been right. The model you would have picked today is not the one a tool vendor picked eighteen months ago when they integrated.

By contrast, a tool that lets you bring your own model is agnostic to the bet. The product gets better when models get better. You get to choose, on your timeline, when to switch providers. The tool's value compounds with the model market, instead of being capped by whichever model the vendor chose.

This is the architecture every serious AI tool will end up with eventually. The teams that adopt it now save themselves the inevitable migration in two years when the locked-in vendor falls behind.

What changes when the model is your choice

There are several practical effects of running the same product against different models. The user-visible ones, in order of how often they come up:

Response style varies. Claude tends to be more verbose and reflective. GPT-4o tends to be more concise. Gemini tends to be faster on certain operations. Grok has its own pacing. The tool's output reflects the model's voice, and you pick the voice that fits how you like to read your tools.

Cost varies. Different providers price tokens differently. Different models within the same provider price differently. A team running heavy automated workflows can choose a faster, cheaper model for high-volume operations and reserve the premium model for the calls that matter most.

Capabilities vary. Some models handle long contexts better. Some handle code better. Some are better at structured output. Choosing the model for the task means you get the right tool for each kind of work, not the lowest common denominator.

Compliance varies. Some teams cannot send data to certain providers for regulatory reasons. With BYO model, those teams can route their work to the providers that meet their requirements, including running everything locally if they need to.

The local-model option

A particular kind of team cares deeply about a particular kind of guarantee: that no data leaves their network. For these teams, no commercial AI provider is acceptable. The data sensitivity is too high, the regulatory environment is too strict, or the trust relationship is not in place.

For these teams, GitChat ships with LM Studio support. LM Studio runs AI models entirely on your own hardware. You point GitChat at your local LM Studio endpoint. Every AI interaction happens on your machine. No cloud provider sees your data. No third-party API ingests your code.

The capability is bounded by your local model's quality, which is improving rapidly but not yet at the level of frontier cloud models. The trade-off is data residency for capability. For the teams that need it, this trade-off is the only acceptable one.

This is the architectural choice that pays off. A tool that supports BYO model can support both the cutting-edge cloud user and the air-gapped enterprise customer with the same product, the same workflow, and the same trust model. The lock-in vendor cannot do this. The vendor's product is only as private as their model partner allows.

The cost story, made explicit

One of the less-discussed effects of bring-your-own-model is what it does to pricing.

Most AI developer tools charge a per-seat subscription that bundles the model cost into the seat. The vendor pays the model provider; you pay the vendor; the markup is invisible. The convenience is that you do not have to think about token costs. The cost is that you pay a flat rate regardless of how much you use, and the markup can be substantial.

With bring-your-own-model, the model cost is yours, paid directly to the provider. You see exactly what you are paying. You can throttle if you want. You can pick a cheaper model for the work that does not need a frontier one. The tool charges you for the tool, not for tokens.

For light users, this is dramatically cheaper. For heavy users, it is roughly comparable but more transparent. For teams with existing model relationships, it is mostly already paid for. In every case, you know what your money is buying.

What we believe about the future

We think the next decade of AI developer tools is going to look different from the current one. The early movers will continue to compete on model partnerships and tight vertical integration. The teams that win, in our view, will be the ones that decouple the product layer from the model layer, the way good software has always decoupled the application layer from the database layer or the storage layer.

GitChat is built on this belief. The product is the conversational interface to your repository. The AI model is the engine that runs the conversation. The two are separable. You can change the engine without changing the interface. You can keep your existing model subscriptions. You can switch providers when the market shifts. You can run locally when you have to.

This is the architecture we picked when we started building, and the architecture we will keep. The model lock-in problem is real. We did not want to be part of it. We want our users to be able to bring their best tool to the best workflow, not the other way around.

If you have read this far, you probably already have a favorite AI model. Use it. We built the product to work with whichever one you pick. That is the architecture we believe in, and the one we hope becomes the default for every serious AI tool that ships from here.