pagent (English)

Language: 中文 | English · Docs · For agents · llms.txt

pagent is a small async Python library for an Agent + tools loop over OpenAI-compatible Chat Completions. Good for scripts, experiments, and teaching—transparent message history, your own tools.

Documentation

https://synclionpaw.github.io/pagent/ — install, quick start, tools, events, Wire, providers.

Install

Requires Python 3.11+.

pip

pip install pagent
pip install "pagent[search]"   # optional web_search tool

uv

uv is a fast Python package and project manager (official docs).

uv pip install pagent
uv pip install "pagent[search]"

# or in a uv-managed project
uv add pagent
uv add "pagent[search]"

uvx (terminal REPL)

export DEEPSEEK_API_KEY="your-key"
uvx pagent
uvx pagent --thread-id demo

conda

conda activate your-env
pip install pagent
pip install "pagent[search]"

Conda envs usually install PyPI packages with pip inside the activated environment. Check conda-forge if you prefer a conda package when available.

Quick start

import asyncio
import os

from pagent import Agent, LLM, Session, tool


@tool()
def get_weather(city: str) -> str:
    """Return a fake weather summary for the city."""
    return f"It's sunny in {city} today."


async def main() -> None:
    if not os.getenv("OPENAI_API_KEY"):
        raise SystemExit("Please set OPENAI_API_KEY first.")

    agent = Agent(
        llm=LLM("gpt-4o-mini"),
        session=Session("You are a concise assistant. Use tools when needed."),
        tools=[get_weather],
        max_turns=24,
    )

    result = await agent.run("What's the weather in Xiamen?")
    print(result.content)
    print(agent.stats)


asyncio.run(main())

run() returns RunEnd; use .content for the answer.

Streaming & events

One agent timeline — pick an API by consumer (not two different event systems):

Wire serializes the same events as native Event; see docs/events.md and docs/wire.md.

Minimal event consumer (build your own UI or logs):

import asyncio

from pagent import (
    Agent,
    LLM,
    RunEnd,
    Session,
    TextDelta,
    ToolCallBegin,
    ToolResult,
)


async def main():
    agent = Agent(LLM("gpt-4o-mini"), Session("You are helpful."), tools=[])

    async for event in agent.arun_events("What is 2 + 2?"):
        if isinstance(event, TextDelta):
            print(event.text, end="", flush=True)
        elif isinstance(event, ToolCallBegin):
            print(f"\n[calling {event.name}]", flush=True)
        elif isinstance(event, ToolResult):
            print(f" {event.content}", flush=True)
        elif isinstance(event, RunEnd):
            print(f"\n\n(done, {event.content!r})")


asyncio.run(main())

Common events: TextDelta (answer stream), ReasoningDelta (model thinking, if supported), ToolCallBegin / ToolResult, RunEnd (final result with .content and .reasoning_content).

Full list: docs/events.md. Frontend / JSON: docs/wire.md — each line is {"jsonrpc":"2.0","method":"TextDelta","params":{...}}.

async for line in agent.arun_wire("Hello"):
    # send `line` over SSE / WebSocket (already ends with \n)
    ...

Runnable demos are grouped under examples/; start with examples/README.md.

Models & API keys

from pagent import DeepSeek, Ollama

llm = DeepSeek("deepseek-v4-flash")
llm = Ollama("llama3.2")

Server must expose OpenAI-compatible /v1/chat/completions.

Examples

Guide: docs/reasoning.md. Full-stack wire demo: examples/wire_browser/.

export DEEPSEEK_API_KEY="your-key"
uv run python -m examples.pagentv4.thread_based.conversation_only

Optional built-in tools

from pagent import Agent, LLM, Session, web_search

agent = Agent(LLM("gpt-4o-mini"), Session("..."), tools=[web_search])

See clock, region in pagent.defaults.

Notes

• Requires an OpenAI Chat Completions–compatible API.
• A minimal embeddable loop—not a full coding-agent product with file edit/shell.
• Development & internals: docs/development.md

License

MIT