v1.0.0 · Apache 2.0 · Go 1.22+
One Go interface for every AI model.
Talk to Claude, GPT, Gemini, Ollama and 400+ other models through a single, stable interface — then switch between them by changing one string.
client := skyl.New(anthropic.New(os.Getenv("ANTHROPIC_API_KEY")))
resp, err := client.Complete(ctx, &skyl.Request{ Model: "claude-opus-5", Messages: []skyl.Message{skyl.UserText("Hello")},})
fmt.Println(resp.Text())
The highlighted line is the only one that changes.
Why skyl #
Integrating an AI model into a Go service means writing the same 400 lines every time: request mapping, SSE parsing, retry with jitter, rate-limit backoff, token accounting, error classification. skyl does that once, properly, with tests.
Provider-agnostic
One Request and Response shape across every vendor. Your application code never learns a vendor’s JSON.
Zero-day model support
Model IDs are pass-through strings. A model released this morning works this morning, with no skyl release.
No lock-in
Every response carries Raw — the untouched provider JSON — so skyl’s abstraction is never the reason you cannot ship.
Streaming that works
A pull iterator with proper context cancellation, verified by a goroutine-leak test rather than by convention.
Honest errors
Typed and classified, so you branch with errors.Is instead of matching on message text that vendors reword.
Small surface
The core module has zero external dependencies. No router, no logger, no framework imposed on you.
Every model, one interface #
Three native adapters for full vendor fidelity, and one generic adapter that reaches the long tail of hosts speaking OpenAI’s wire format.
provider/anthropic own module
Claude, native. A separate module, because it is built on the official Anthropic SDK and that brings a dozen transitive dependencies nobody else should pay for.
Claude Opus 5, Fable 5, Sonnet 5, Haiku 4.5, and the 4.x family
provider/openai
GPT, native. Ships inside the core module, so it costs you no dependencies.
GPT-5.6 (Sol, Terra, Luna), GPT-5.5, GPT-5.4 nano, the o-series, and gpt-oss
provider/gemini
Gemini, native. Handles Gemini’s structural differences — contents/parts, the "model" role in place of "assistant", and systemInstruction — so your Request does not have to.
Gemini 3.6 Flash, 3.5/3.1 Flash-Lite, and 3 Pro
provider/openaicompat
One adapter, configured per host, reaching the long tail of vendors that speak OpenAI's wire format — including local runtimes that need no credential at all.
~18 hosts including xAI, DeepSeek, Groq, OpenRouter, Ollama, vLLM and LM Studio
18 verified compatible endpoints — including Ollama, vLLM and LM Studio, which need no credential at all.
Try it without an API key #
skyl ships a sandbox that speaks all four providers’ wire protocols locally. Every example in this documentation runs against it — no signup, no credential, no cost.
go run ./cmd/skyl-sandbox
Status: v1.0.0 — the API is stable, the feature surface is not complete #
Every adapter has been exercised against its live provider API, so the wire mapping is confirmed rather than merely self-consistent. Since v1.0.0 the exported API is frozen: no breaking change without a major version. What is not promised is complete provider coverage. Each adapter still silently ignores some provider features, and embeddings, prompt-caching control, batch APIs, token counting and failover are deliberately deferred. Both lists are published rather than implied.