Python LLM Client

Install brokk-anvil-client with uv add brokk-anvil-client. This is a native PyO3 binding to Anvil’s standalone client, separate from the brokk-anvil CLI launcher. Wheels do not require Rust or the Anvil executable. Standard CPython 3.11+ is supported on Linux x86-64/ARM64, macOS ARM64, and Windows x86-64.

from anvil_client import Client

async def classify(text: str):
    async with Client() as client:
        result = await client.infer(
            model="deepseek::deepseek-v4-flash",
            messages=[{"role": "user", "content": text}],
            schema={"type": "object", "properties": {"relevant": {"type": "boolean"}},
                    "required": ["relevant"], "additionalProperties": False},
        )
        return result.output

The client accepts the same explicit hosted-provider routes and credentials as anvil infer: Codex, Meta, Kimi, Grok, and DeepSeek. It reuses native authentication, retries, schema validation, and usage accounting. No tools, agent sessions, or project instructions run.

DeepSeek structured inference uses its stateless Responses API with native text.format: json_schema requests. DeepSeek can still return schema violations even with strict: true; the client validates locally and uses bounded repair retries before returning an error. Its beta strict tool-call mode also cannot be relied on to enforce nested schemas. Truncated responses are rejected even when their partial text happens to be valid JSON. The schema is sent out of band, so changing it does not prepend instructions ahead of your stable message prefix. Other providers that require JSON-mode fallback receive the schema after your messages; local validation checks every provider’s output. DeepSeek’s normal agent chat continues to use Chat Completions.

Keep one client for repeated calls to reuse provider connections. Use it as an async context manager or call close(); closing cancels outstanding requests. Asyncio task cancellation cancels native inference. Credentials are resolved lazily and retained for the client’s lifetime, so create a new client after changing credentials.

infer() takes model, system/user messages, schema, and optional schema_name, reasoning_effort, service_tier, timeout (120 seconds), idle_timeout (60), stall_timeout (30), and validation_retries (1). Timeouts must be finite and positive. Results provide output, usage, model, reasoning_effort, and service_tier. Exceptions derive from InferenceError and expose kind, with specific subclasses for invalid input, authentication, rate limits, context length, transport, schema validation, and deadlines.

Maintainers build this package from crates/anvil-client-python using Maturin. The package is versioned with Anvil. The Python client wheels workflow checks installed wheels on Python 3.11 and 3.14, publishes tagged wheels as GitHub release assets. A manual workflow dispatch with publish_pypi enabled publishes to PyPI using the pypi-publish environment. For the initial project bootstrap, a maintainer can publish downloaded, validated wheels with an existing PyPI token; do not store that token in the repository. Configure a PyPI trusted publisher for project brokk-anvil-client, owner BrokkAi, repository anvil, workflow python-client.yml, environment pypi-publish before its first publication. Wheels carry Anvil’s LGPL license; the tagged repository is their corresponding source.