Why we built ByteSpike
Every team we know is gluing two or three model providers together with bash scripts. We got tired of it. So we built one gateway, one key, and one billing surface — and we made it Anthropic-compatible because that's the API we actually like writing.
Halfway through 2025 we counted: between Claude, GPT-4o, Gemini, DeepSeek, Doubao, and the various image and video endpoints, our internal projects were talking to nine different providers. Each one had its own SDK, its own retry semantics, its own error vocabulary, its own way of pricing failed requests. Every new model launch meant another integration weekend.
We're a small team. We don't want to be a switching layer. We want to ship product. So we wrote down what would actually solve this for us, and ByteSpike is the answer.
What ByteSpike actually is
One gateway. Two protocols (Anthropic Messages and OpenAI Chat Completions). Every frontier model behind one Anthropic-style key. Failures don't bill. Every response carries an estimated_credits header so you can show the cost before the user confirms.
- Anthropic-compatible by default — keep your tool_use, cache_control, and thinking blocks.
- OpenAI Chat Completions transparently shimmed for the long tail of existing code.
- Multimodal endpoints (image, video, audio, embedding, rerank) live behind the same key.
- Org-level wallets and per-member quotas, sub-keys derived from a master key.
What we deliberately didn't build
We didn't build a chat UI. There are enough chat UIs. If you want one, DOSIA is our companion app — same account, native macOS, opinionated, and quietly excellent. We didn't build a vendor-neutral abstraction either; calling tool_use a 'function' just to feel generic is how you lose fidelity to the providers that actually ship.
“Pick the API of the company that ships fastest, then make every other model speak that API.”
What we owe the open-source world
ByteSpike's gateway is built on top of sub2api (LGPLv3) by Wei-Shaw. We've kept the upstream attribution in the footer because we should, and we'll be contributing the multimodal routing and pool work back upstream over the next few releases.
What's next
Self-hosted SKU is in private preview with two enterprise design partners. Mintlify-powered docs at docs.bytespike.ai land alongside V1.0. After that, we'll start shipping vertical recipes — RAG over a Lark base, screenshot-to-tickets, the things our own marketing and ops teams already use DOSIA for every day.
If any of this resonates, the playground is three free runs per IP, no signup. We'd love to hear what we're missing.