Getting started¶

Install¶

pip install "glia-agents[anthropic]"   # core + Claude provider

For local development from a clone:

git clone https://github.com/DenisDrobyshev/glia
cd glia
pip install -e ".[anthropic,dev]"

The distribution name is glia-agents; the import is always import glia.

Your first agent¶

import asyncio
from glia import Agent, tool
from glia.providers import ClaudeLLM

@tool
async def add(a: int, b: int) -> int:
    """Add two integers."""
    return a + b

async def main():
    agent = Agent(ClaudeLLM(), tools=[add], system="You are precise.")
    result = await agent.run("What is 21 + 21?")
    print(result.output)

asyncio.run(main())

ClaudeLLM defaults to the claude-opus-4-8 model and resolves credentials the same way the Anthropic SDK does (ANTHROPIC_API_KEY, or an ant auth login profile).

Run it offline — no API key¶

glia ships a deterministic provider, EchoLLM, that replays a scripted sequence of turns. It satisfies the same LLM protocol as ClaudeLLM, so your agent code never changes:

from glia.providers import EchoLLM, call

# Turn 1: the model calls the tool. Turn 2: it answers.
llm = EchoLLM([call("add", {"a": 21, "b": 21}), "The answer is 42."])
agent = Agent(llm, tools=[add])
result = await agent.run("What is 21 + 21?")
assert result.output == "The answer is 42."

This is what makes glia testable: the entire loop — tools, subagents, compaction, approval — runs in CI with no cost and no flakiness.

Watch every step¶

The agent emits an event for everything it does. Consume the event stream to see the glass box working:

async for event in agent.run_events("What is 21 + 21?"):
    print(event.kind)
# run_started → model_call → model_response → tool_called → tool_returned → ... → run_finished

Turn on streaming¶

agent = Agent(ClaudeLLM(), tools=[add], stream=True)
async for event in agent.run_events("Explain addition briefly."):
    if event.kind == "model_delta":
        print(event.text, end="", flush=True)   # tokens as they arrive

Where to next¶

The Guide covers every primitive with examples.
The examples/ folder in the repo has seven runnable scripts (all offline).