glia

English · Русский · Docs

A glass-box, minimal library for building LLM agents. Every model call, tool call, and state transition is a plain object you can log, snapshot, and replay. No hidden control flow. The whole loop fits in one file you can read in an afternoon.

glia (n.): the cells that support and connect neurons. This is the connective tissue for LLM agents — not a framework you submit to, a small library you build on.

 Python 3.10+  ·  MIT  ·  zero required dependencies  ·  typed

_{The desktop shell — streaming reply on the left, the live glass box on the right. (Animated SVG; it plays in your browser.)}

Why another agent library?

The 2026 agent-framework field is crowded, and the loudest, most consistent complaint about the incumbents is the same: too much abstraction, hidden control flow, painful to debug. Developers keep stripping the framework out to call the model API directly, just to see what's happening.

glia is the opposite bet. It ships the modern techniques — tools, structured outputs, context compaction, durable checkpoints, guardrails, subagents, evals-as-tests — as opt-in primitives you can read, not a monolith you must trust. The design goal is understandability and control, not feature count.

If you want a graph engine, use LangGraph. If you want role-play crews, use CrewAI. If you want a small, transparent loop you fully understand — glia.

See docs/STRATEGY.md for the full market analysis.

Install

pip install glia-agents               # core — no dependencies
pip install "glia-agents[anthropic]"  # + the Claude provider

The distribution is glia-agents (the bare name glia was taken on PyPI); the import stays import glia. For development: git clone then pip install -e ".[anthropic,dev]".

30-second tour

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

@tool
async def get_weather(city: str) -> str:
    """Get the current weather for a city."""
    return {"Paris": "18°C, cloudy"}.get(city, "unknown")

async def main():
    agent = Agent(ClaudeLLM(), tools=[get_weather], system="Be concise.")
    result = await agent.run("What's the weather in Paris?")
    print(result.output)          # the answer
    print(result.usage)           # what it cost

asyncio.run(main())

No API key handy? Every example runs offline with the deterministic EchoLLM provider — same code, no network:

from glia.providers import EchoLLM, call
llm = EchoLLM([call("get_weather", {"city": "Paris"}), "It's 18°C and cloudy."])

See the whole glass box

Because the loop emits an event for everything it does, you can watch it work:

async for event in agent.run_events("What's the weather in Paris?"):
    print(event.kind)
# run_started → model_call → model_response → tool_called → tool_returned → model_call → ... → run_finished

And the entire run state is one serialisable object:

from glia.checkpoint import save, load
save(result.trajectory, "run.json")     # durable execution: it's just JSON
resumed = load("run.json")
await agent.run("follow-up question", trajectory=resumed)   # pick up where you left off

What's in the box

Examples

Runnable, offline, no API key needed:

python examples/01_hello_agent.py       # basic agent + event stream
python examples/02_tools.py             # tools
python examples/03_structured_output.py # typed output
python examples/04_subagents.py         # subagent as a tool
python examples/05_checkpoint_resume.py # durable execution
python examples/06_evals.py             # eval suite
python examples/07_streaming_and_approval.py # streaming + parallel tools + approval gate
python examples/08_ollama_local.py           # run a local model (Qwen/DeepSeek) via Ollama
python examples/09_record_replay.py          # record once, replay deterministically offline

Desktop app

glia ships a graphical shell — a chat window that also shows the live glass box (streaming tokens, tool calls, approvals) as it happens.

• Download a standalone build for your OS from the latest release (glia-shell-windows.exe / -macos / -linux) — no Python needed.
• Or pip install "glia-agents[shell]" then glia-shell for a native window.

It works offline in demo mode out of the box; add an Anthropic API key in Settings to chat with real Claude. The whole app is stdlib + one HTML file — see docs/app.md.

Design in one picture

Agent.run() → [ call model → (tools? run them, loop) : done ]
                     │              │
                every step emits an Event you can see, log, and replay
                     │
              all state lives in one serialisable Trajectory

Full details in docs/ARCHITECTURE.md.

Project docs

• Documentation site — full guide, available in English and Русский
• Strategy & market analysis — why glia, and who it competes with
• Architecture — how the whole thing works
• Roadmap — where it's going
• Contributing

Status

v0.8 — alpha. The core thesis is proven end-to-end with a full test suite (167 offline tests, ~95% coverage) and green CI (ruff + mypy + tests). v0.8 redesigns the desktop shell into a modern multi-conversation app (sidebar with persistent history, Markdown/code rendering, theme toggle); v0.7 an OpenTelemetry exporter; v0.6 record/replay cassettes; v0.5 OpenAI + interactive approval + MCP; v0.4 local models via Ollama; v0.3 a downloadable desktop shell; v0.2 streaming, parallel tools, and the approval gate. APIs may still change before 1.0. Feedback and issues welcome.

License

MIT — see LICENSE.