Native targets: Linux, Windows, macOS

Ambi's native path supports OpenAI-compatible APIs, local llama.cpp inference and custom engines. The platform-specific choice should stay outside your higher-level Pipeline whenever possible.

Runtime and build

[dependencies]
ambi = "0.4.1"
tokio = { version = "1", features = ["rt-multi-thread", "macros"] }

The recommended native setup uses Tokio's multi-thread runtime because Agent::make() routes engine initialization through the crate's blocking-task adapter. Do not rely on a current_thread setup without testing your exact backend and executor.

To build local inference:

ambi = { version = "0.4.1", default-features = false, features = ["llama-cpp"] }

GPU-specific features are cuda, vulkan, metal, rocm; do not enable more than one. Hardware tools, SDKs and driver requirements depend on the target platform. dynamic-link is available where appropriate.

Use the conservative LlamaEngineConfig::builder() defaults and override context, batch size, model path and GPU settings as needed. Local inference runs behind a worker queue; session switching may require KV re-prefill.

Smoke tests

cargo check --no-default-features --features openai-api
cargo test --no-default-features --features openai-api
cargo test --no-default-features --features "openai-api,sub-agent" --test leaf_subagent_test

The last command exercises the optional Leaf integration tests and requires a compatible environment. It is not included in every project's default test invocation automatically.

See Getting started and Architecture.