lm15-dev

Documentation · Playground · Python · TypeScript · Rust · Go · Contract

lm15 lets a program talk to OpenAI, Anthropic, Google Gemini, xAI, Groq, DeepSeek, OpenRouter, Z.AI, Moonshot, Meta, the clouds (Azure, AWS Bedrock, Google Vertex) and models on your own machine through one set of types. Write a request once; change the model string to change the provider.

from lm15 import LMRouter, Message, Request

router = LMRouter()   # API keys come from the environment
response = router.complete(Request(
    model="anthropic:claude-haiku-4-5",   # or "gpt-4.1-mini", "gemini:gemini-2.5-flash", "ollama:qwen3.5:0.8b"
    messages=(Message.user("What eats acorns at night?"),),
))
print(response.text)

It is a foundation library: typed requests, responses, stream events, tools, media, errors and exact JSON, built on each language's standard library. No hidden tool loop, no retries you didn't ask for, no prompt templates. It is the layer you build your own opinions on.

All four released languages pass every check of the shared contract at the version they pin: the same program builds the same request and reads the same answer from the same reply in every language. Early ports in Java, Ruby, Swift and .NET were written against an earlier contract and are not published.

Four nouns, and one more for streaming:

Part  →  Message  →  Request  ⇢  Response
                        ⇣ (streaming)
              start → Delta… → end

• Part: the atom of content, one of twelve kinds: text, image, audio, video, document, binary, tool_call, tool_result, thinking, refusal, citation, data.
• Message: a role (user, assistant, developer, tool) and parts.
• Request: a model, messages, and optionally system, tools and a config (length, temperature, reasoning, caching, tool choice, structured output…).
• Response: an assistant message, a finish_reason from a closed list, and usage, where a missing number means "the provider didn't say", never a silent 0.
• Delta: while streaming, typed fragments between exactly one start and one end event. They assemble into the same Response a non-streamed call returns.

Provider-specific settings go in through extensions and provider-specific data comes out through provider_data, both passed through untouched. When a provider can't do what a request asks, lm15 either adapts and records what it changed, or refuses before sending: never a silent drop.

Direct APIs: OpenAI, Anthropic, Google Gemini (with Live sessions), xAI, Groq, DeepSeek, OpenRouter, Z.AI, Moonshot / Kimi, Meta, TypeSafe. Clouds: Azure (OpenAI and Anthropic), AWS Bedrock, Google Vertex AI. Local and self-hosted: Ollama, vLLM, SGLang and any server that speaks OpenAI Chat Completions. Accounts: ChatGPT (Codex), Claude, xAI, GitHub Copilot, Kimi Code and OpenRouter sign-in.

Beyond chat: streaming, function and built-in tools, structured output, judgments with probabilities, reasoning controls, prompt caching, files, batches, image and speech generation, video, realtime sessions, the model catalog, and reading an OpenAI Chat Completions request into lm15.

lm15-contract is the single source of truth; no implementation is.

• A written specification: every type, field, default and validation rule, 53 numbered invariants, 16 mapping rules, and the rules for exact JSON.
• Recorded provider traffic: real requests and replies, each with a receipt of when and against which model it was captured.
• A language-neutral harness that grades an implementation in 18 directions (requests, responses, streams, errors, serialization, credentials, models, files, batches, caches, live sessions, sign-in…) and never trusts the implementation under test.
• Evidence rules enforced by CI: recorded traffic changes only with a new live capture; the specification changes only with a written decision.

Method and full results: BENCHMARKS.md.