Yeah I share the same uncertainty here. My understanding is personal and interactive use should be fine. I use Conductor all day every day and it wraps a subscription.
Perhaps fully automated use is where the line is drawn.
But I also suspect individuals using it for light automated dispatching would be ok too.
i use this for my personal projects. some features are gated behind a license but the basics like provider proxy, logs, metrics are covered in the free version.
https://github.com/maximhq/bifrost
> Any other Go-based AI / LLM tools folks are happy with?
I can throw my hat into the ring, built on ADK, CUE, and Dagger (all also in Go); CLI, TUI, and VSCode interfaces. It's my personal / custom stack, still need to write up docs. My favorite features are powered by Dagger, sandbox with time travel, forking, jump into shell at any turn, diff between any points.
Hey, this looks super nice. I do like the 'compact' feel of this. Reminds me of Traefik. It seems very promising indeed!
One problem I have is that yes, LiteLLM key creation is easier than creating it directly at the providers and managing it there for team members and test environments, but if I had a way of generating keys via vault, it would be perfect and such a relief in many ways.
I see what I need on your roadmap, but miss integration with service where I can inspect and debug completion traffic, and I don't see if I would be able to track usage from individual end-users through a header.
"... and I don't see if I would be able to track usage from individual end-users through a header".
Currently we have a unified concept of User-Paths. Once you add a specific header OR assign User-Path to an API key, you can track the usage based on this. The User-Path might be youe end-user, internal user or some service. Examples:
Given this app seems to expose itself via REST calls, why would anyone care that it’s written in Go? I guess it matters to potential contributors but the majority of interest would be from users.
It's like fuel costs in a supply chain. When you buy apples at the store, you don't think about oil prices. But if trucks ran on something cheaper, more efficient, or less taxed, the apples on the shelf would be cheaper too.
I have written and maintained AI proxies. They are not terribly complex except the inconsistent structure of input and output that changes on each model and provider release. I figure that if there is a not a < 24 hour turn around for new model integration the project is not properly maintained.
Governance is the biggest concern at this point - with proper logging, and integration to 3rd party services that provide inspection and DLP type threat mitigation.
Another reason golang is interesting for the gateway is having clear control of the supply chain at compile time. Tools like LiteLLM the supply chain attacks can have more impact at runtime, where the compiled binary helps.
Does this have a unified API? In playing around with some of these, including unified libraries to work with various providers, I've found you are, at some point, still forced to do provider-specific works for things such as setting temperatures, setting reasoning effort, setting tool choice modes, etc.
What I'd like is for a proxy or library to provide a truly unified API where it will really let me integrate once and then never have to bother with provider quirks myself.
Also, are you also planning on doing an open-source rug pull like so many projects out there, including litellm?
Are these kinds of libraries a temporary phenomenon? It strikes me as weird that providers haven't settled on a single API by now. Of course they aren't interested in making it easier for customers to switch away from them, but if a proprietary API was a critical part of your business plan, you probably weren't going to make it anyway.
(I'm asking only about the compatibility layer; the other tracking features would be useful even if there were only one cloud LLM API.)
I've been maintaining an abstraction layer over multiple providers for a couple of years now - https://llm.datasette.io/
The best effort we have to defining a standard is OpenAI harmony/responses - https://developers.openai.com/cookbook/articles/openai-harmo... - but it's not seen much pickup. The older OpenAI Chat Completions thing is much more of an ad-hoc standard - almost every provider ends up serving up a clone of that, albeit with frustrating differences because there's no formal spec to work against.
The key problem is that providers are still inventing new stuff, so committing to a standard doesn't work for them because it may not cover the next set of features.
2025 was particularly turbulent because everyone was adding reasoning mechanisms to their APIs in subtly different shapes. Tool calls and response schemas (which are confusingly not always the same thing) have also had a lot of variance - some providers allow for multiple tool calls in the same response, for example.
My hunch is we'll need abstraction layers for quite a while longer, because the shape of these APIs is still too frothy to support a standard that everyone can get behind without restricting their options for future products too much.
The providers themselves can't keep this straight even within their own ecosystem. Plus everyone is running at a million miles/hour.
For example `Claude code` used to set 2 specific beta headers with some version numbers for their Max subscription to be supported.
Oauth tokens for Max plan is different from how their API keys looked. They kind of look similar, but has specific prefix that these tool pre-validate.
It is barely working at this point even within a single provider
It’s a complete mess, and the hardest part of this kind of tool is maintenance.
It’s not just about incompatible APIs, but also about how messages are structured. Even getting reliable tool calling requires a significant amount of work and testing for each individual model.
Just look at LiteLLM’s commit history and open issues/PRs. They’re still struggling with reliable multi-turn tool calling for Gemini, Kimi requires hardcoded rules (so K2.6 is currently unsupported because it’s not on the list), and so on.
Implementing the basic, generic OpenAI/Anthropic protocols is trivial, and at that point it almost feels like building an AI gateway is done. But it isn’t — that’s just the beginning of a long journey of constantly dealing with bugs, changes, and the quirks of each provider and model.
How do you plan on keeping up with upstream changes from the API providers? I have implemented something similar, and the biggest issue I have faced with go is that providers don’t usually have sdk’s (compared to javascript and python), and there is work involved in staying up to date at each release.
First, GoModel is designed to be flexible. If you add an extra field, it tries to pass it through in the appropriate place (Postel's law)
Therefore there's a good chance that if they make a minor API-level change, GoModel will handle it without any code changes.
Also, changes to providers' API formats might be less and less frequent. Keeping up typically means adding a few lines of code per month. I'm usually aware of those changes because I use LLMs daily and follow the news in a few places.
As a fallback, GoModel includes a passthrough API that forwards your request to the provider in its original format. That might be useful when an AI provider changes their contract significantly and we haven't caught up yet.
Also, official SDKs aren't bug-free either. Skipping that extra layer and hitting the API directly might actually be beneficial for GoModel.
LiteLLM proxy has to be about the worst codebase and performing Python code I've used in a long time. Incredibly poor performance, riddled with serious bugs and to be honest devs that don't seem to understand them.
Curious how the semantic caching layer works.. are you embedding requests on the gateway side and doing a vector similarity lookup before proxying? And if so, how do you handle cache invalidation when the underlying model changes or gets updated?
Hey, contributor here. That's right, GoModel embeds requests and does vector similarity lookup before proxying. Regarding the cache invalidation, there is no "purging" involved – the model is part of the namespace (params_hash includes the LLM model, path, guardrails hash, etc). TTL takes care of the cleanup later.
This is way more interesting to me as well. I have projects that use small limited-purpose language models that run on local network servers and something like this project would be a lot simpler than manually configuring API clients for each model in each project.
You are not the first person who has asked about it.
It looks like a useful feature to have. Therefore, I'll dig into this topic more broadly over the next few days and let you know here whether, and possibly when, we plan to add it.
This is really useful. I've been building an AI platform (HOCKS AI) where I route different tasks to different providers — free OpenRouter models for chat/code gen, Gemini for vision tasks. The biggest pain point has been exactly what you describe: switching models without changing app code.
One thing I'd love to see is built-in cost tracking per model/route. When you're mixing free and paid models, knowing exactly where your spend goes is critical. Do you have plans for that in the dashboard?
It’s a heavily vibe coded project with only proxy with terrible benchmarks design. Basically vibe coded benchmarks that lie through ignorance of mocked super fast endpoint without using full power of litellm in multiple processes.
Other than that almost useless it’s faster when this will be io bound and not cpu bound.
GoModel. I see some red flags in the docs/benchmarks, but I could be wrong in my judgement here.
What I noticed: the website shows a diagram of the litellm SDK communicating with the gateway proxy of GoModel, poor design of benchmarks, the scope of the project in readme vs. depth.
I don't have professional experience in GoLang, so will not comment on quality of code.
There are some genuinely good things about this project and the effort here, but with solid position of Bifrost sitting at a version above 1.0.0 and so many other initiatives in this space, it's a tough market.
If your concern is someone selling GoModel as a service, you could add a license provision for that. Technically it'd no longer be open source, I think, but most people won't care.
I'm all in on Go and integrating AI up and down our systems for https://housecat.com/ and am currently familiar and happy with:
https://github.com/boldsoftware/shelley -- full Go-based coding agent with LLM gateway.
https://github.com/maragudk/gai -- provides Go interfaces around Anthropic / OpenAI / Google.
Adding this to the list as well as bifrost to look into.
Any other Go-based AI / LLM tools folks are happy with?
I'll second the request to add support for harnesses with subscriptions, specifically Claude Code, into the mix.
It's a just-bash like variant implemented in Go. Useful for giving a managed bash tool to your agents without a full sandboxing solution.
However, it might be challenging, considering that Claude Code with a subscription no longer officially works with OpenClaw.
Perhaps fully automated use is where the line is drawn.
But I also suspect individuals using it for light automated dispatching would be ok too.
https://news.ycombinator.com/item?id=47844269
I can throw my hat into the ring, built on ADK, CUE, and Dagger (all also in Go); CLI, TUI, and VSCode interfaces. It's my personal / custom stack, still need to write up docs. My favorite features are powered by Dagger, sandbox with time travel, forking, jump into shell at any turn, diff between any points.
Good entrypoint folder: https://github.com/hofstadter-io/hof/tree/_next/lib/agent
One problem I have is that yes, LiteLLM key creation is easier than creating it directly at the providers and managing it there for team members and test environments, but if I had a way of generating keys via vault, it would be perfect and such a relief in many ways.
I see what I need on your roadmap, but miss integration with service where I can inspect and debug completion traffic, and I don't see if I would be able to track usage from individual end-users through a header.
Thank you and godspeed!
Currently we have a unified concept of User-Paths. Once you add a specific header OR assign User-Path to an API key, you can track the usage based on this. The User-Path might be youe end-user, internal user or some service. Examples:
Would this work for you?https://gomodel.enterpilot.io/docs/features/user-path
PS Thanks for the feedback on the Vault integration. Noted.
Governance is the biggest concern at this point - with proper logging, and integration to 3rd party services that provide inspection and DLP type threat mitigation.
https://sbproxy.dev - engine is fully open source.
Another reason golang is interesting for the gateway is having clear control of the supply chain at compile time. Tools like LiteLLM the supply chain attacks can have more impact at runtime, where the compiled binary helps.
What I'd like is for a proxy or library to provide a truly unified API where it will really let me integrate once and then never have to bother with provider quirks myself.
Also, are you also planning on doing an open-source rug pull like so many projects out there, including litellm?
2. Regarding being open-source and the license, I've described our approach here transparently: https://gomodel.enterpilot.io/docs/about/license
(I'm asking only about the compatibility layer; the other tracking features would be useful even if there were only one cloud LLM API.)
The best effort we have to defining a standard is OpenAI harmony/responses - https://developers.openai.com/cookbook/articles/openai-harmo... - but it's not seen much pickup. The older OpenAI Chat Completions thing is much more of an ad-hoc standard - almost every provider ends up serving up a clone of that, albeit with frustrating differences because there's no formal spec to work against.
The key problem is that providers are still inventing new stuff, so committing to a standard doesn't work for them because it may not cover the next set of features.
2025 was particularly turbulent because everyone was adding reasoning mechanisms to their APIs in subtly different shapes. Tool calls and response schemas (which are confusingly not always the same thing) have also had a lot of variance - some providers allow for multiple tool calls in the same response, for example.
My hunch is we'll need abstraction layers for quite a while longer, because the shape of these APIs is still too frothy to support a standard that everyone can get behind without restricting their options for future products too much.
For example `Claude code` used to set 2 specific beta headers with some version numbers for their Max subscription to be supported.
Oauth tokens for Max plan is different from how their API keys looked. They kind of look similar, but has specific prefix that these tool pre-validate.
It is barely working at this point even within a single provider
It’s not just about incompatible APIs, but also about how messages are structured. Even getting reliable tool calling requires a significant amount of work and testing for each individual model.
Just look at LiteLLM’s commit history and open issues/PRs. They’re still struggling with reliable multi-turn tool calling for Gemini, Kimi requires hardcoded rules (so K2.6 is currently unsupported because it’s not on the list), and so on.
Implementing the basic, generic OpenAI/Anthropic protocols is trivial, and at that point it almost feels like building an AI gateway is done. But it isn’t — that’s just the beginning of a long journey of constantly dealing with bugs, changes, and the quirks of each provider and model.
How do you plan on keeping up with upstream changes from the API providers? I have implemented something similar, and the biggest issue I have faced with go is that providers don’t usually have sdk’s (compared to javascript and python), and there is work involved in staying up to date at each release.
Therefore there's a good chance that if they make a minor API-level change, GoModel will handle it without any code changes.
Also, changes to providers' API formats might be less and less frequent. Keeping up typically means adding a few lines of code per month. I'm usually aware of those changes because I use LLMs daily and follow the news in a few places.
As a fallback, GoModel includes a passthrough API that forwards your request to the provider in its original format. That might be useful when an AI provider changes their contract significantly and we haven't caught up yet.
Also, official SDKs aren't bug-free either. Skipping that extra layer and hitting the API directly might actually be beneficial for GoModel.
However kudos for the project, we need more alternatives in compiled languages.
It looks like a useful feature to have. Therefore, I'll dig into this topic more broadly over the next few days and let you know here whether, and possibly when, we plan to add it.
One thing I'd love to see is built-in cost tracking per model/route. When you're mixing free and paid models, knowing exactly where your spend goes is critical. Do you have plans for that in the dashboard?
However IIUC what you're asking for - it's already in the dashboard! Check the Usage page.
Are there even any benchmarks?
Other than that almost useless it’s faster when this will be io bound and not cpu bound.
What I noticed: the website shows a diagram of the litellm SDK communicating with the gateway proxy of GoModel, poor design of benchmarks, the scope of the project in readme vs. depth.
I don't have professional experience in GoLang, so will not comment on quality of code.
There are some genuinely good things about this project and the effort here, but with solid position of Bifrost sitting at a version above 1.0.0 and so many other initiatives in this space, it's a tough market.
You can use it like this:
It's more lightweight and simpler. The Bifrost docker image looks 4x larger, at least for now.
IMO GoModel is more convenient for debugging and for seeing how your request flows through different layers of AI Gateways in the Audit Logs.