lm15-dev/lm15-python

lm15

lm15 is a small, typed, provider-neutral interface for foundation-model requests, responses, streams, tools, media parts, endpoint APIs, errors, and canonical JSON serialization. This repository is its Python reference implementation.

What lm15 is — and deliberately is not. lm15 is a low-level foundation library: one canonical representation, exact serde for it, and adapters that translate it to and from each provider's wire format — stdlib-only, with its own HTTP transport (websockets is the single optional extra, for live sessions). It is NOT an opinionated user-facing API: no magic call(), no automatic tool loops, no DSL. lm15 is meant to be the dependency for libraries that want to build their own take on the right way to talk to AI systems in Python — you bring the opinions, lm15 brings every provider.

The public API is the top-level package: from lm15 import AnthropicLM, Request, Message, ... (see lm15/__init__.py for the full curated surface). Transport plumbing stays under lm15.transports, live sessions under lm15.live, and the conformance shim under lm15.vet.

The code blocks below are documentation that runs: every output block is the real, captured output of the example above it.

<!-- footprint:generated:start -->

Footprint

Measured by benchmarks/suite/run.py on Python 3.13.3 — methodology and full results in benchmarks/BENCHMARKS.md:

| package | install size | transitive deps | cold import | import RSS | |---|---:|---:|---:|---:| | lm15 | 0.5 MiB | 0 | 152 ms | 16.6 MiB | | openai | 18.0 MiB | 15 | 468 ms | 35.3 MiB | | anthropic | 17.1 MiB | 15 | 589 ms | 41.2 MiB | | google-genai | 37.2 MiB | 24 | 934 ms | 60.8 MiB | | litellm | 133.0 MiB | 54 | 2298 ms | 161.0 MiB | | langchain-openai | 63.3 MiB | 35 | 930 ms | 61.0 MiB |

<!-- footprint:generated:end -->

Install

python3 -m pip install lm15
# Optional extra for websocket live sessio