An agentic toolset.

Reusable, provider-neutral building blocks for agents on low-level LLM SDKs: the six core file/shell tools (Read, Write, Edit, Bash, Glob, Grep), a never-raising dispatch, structured JSONL logging, and a thin adapter per provider. koina gives you the tools and the dispatch; the agentic loop stays in your code.

Requirements

• Python 3.12+
• ripgrep (rg) on PATH (for Glob and Grep)

Install

uv add koina

The library depends only on pydantic. Provider SDKs (anthropic, openai) are the caller's dependency, used in your loop, not by koina.

Usage

dispatch and the tools are provider-neutral; an adapter translates a provider's wire format to and from the neutral ToolCall/ToolResult. With the Anthropic adapter:

from pathlib import Path
from anthropic import AsyncAnthropic
from koina import default_registry, dispatch, ToolContext
from koina.adapters import anthropic as adapter

client = AsyncAnthropic()
reg = default_registry()
ctx = ToolContext(cwd=Path.cwd())
msgs = [{"role": "user", "content": "List the Python files."}]

while True:
    resp = await client.messages.create(
        model="claude-opus-4-8", max_tokens=4096,
        messages=msgs, tools=adapter.tools_param(reg),
    )
    msgs.append({"role": "assistant", "content": resp.content})
    calls = adapter.parse_tool_calls(resp.content)
    if not calls:
        break
    results = [await dispatch(c, reg, ctx) for c in calls]
    msgs.append(adapter.format_results(results))

Swap koina.adapters.anthropic for koina.adapters.openai to run the same tools against the OpenAI Chat Completions API (or any OpenAI-compatible server, e.g. llama.cpp). See examples/ for runnable read-only code-review scripts on both.

What's in the box

• Six core tools (Read, Write, Edit, Bash, Glob, Grep), faithful to Claude Code's observable behavior, headless (no permissions or hooks).
• dispatch never raises: it always returns a ToolResult (errors set is_error=True).
• Provider-neutral core (ToolCall, ToolResult) with per-provider adapters (koina.adapters.anthropic, koina.adapters.openai). The library never imports a provider SDK at runtime.
• Structured logging: typed events (tool calls, model calls, token usage, reasoning) emitted to a pluggable EventSink (JsonlSink/NullSink), so a run reconstructs from a JSONL transcript. Off by default, near-zero overhead when inactive.

Permissions, web tools, and concurrency orchestration are out of scope.