openagentic-sdk 0.1.0

Open-source agent SDK (pure Python)

OpenAgentic SDK (Python)

Pure-Python, open-source Agent SDK inspired by the Claude Agent SDK programming model.

Status: early (APIs may change), but the core runtime + tool loop are usable today.

This project exists for people who want the “agent runtime” experience (multi-turn sessions, tool calls, approvals, skills/commands from .claude/, resumable logs) in a small, hackable Python codebase.

See README.zh_cn.md for a Chinese overview.

What you get

• A minimal agent runtime: run() / streaming query() / CAS-style query_messages().
• A persistent session model: durable session_id, events.jsonl, resume=<session_id>.
• A real tool loop: model requests tools → permission gate → tool execution → tool results → model continues.
• Human-friendly console output by default (debug mode available).
• .claude compatibility: project memory, slash commands, and skills on disk.
• OpenAI + OpenAI-compatible providers (the examples use a real OpenAI-compatible backend by default).

Quickstart (local)

Prereqs: Python 3.11+.

Install (optional, for editable dev):

pip install -e .

Set env (examples + CLI default to RIGHTCODE):

• RIGHTCODE_API_KEY (required)
• RIGHTCODE_BASE_URL (optional, default https://www.right.codes/codex/v1)
• RIGHTCODE_MODEL (optional, default gpt-5.2)

Run unit tests:

python3 -m unittest -q

Run examples:

• python3 example/01_run_basic.py
• See example/README.md for the full list and required env vars.

oa CLI

Install (editable):

pip install -e .

If oa isn't found after installation on Windows, add the scripts directory printed by pip to PATH (or run python -m openagentic_cli chat).

Optional (recommended): install ripgrep (rg) so the agent can search your repo quickly when using shell tools.

• Windows (PowerShell): winget install BurntSushi.ripgrep.MSVC
• WSL/Ubuntu: sudo apt-get update && sudo apt-get install -y ripgrep

Commands:

• oa chat (multi-turn REPL; /help for slash commands)
• oa run "prompt" (--json, --no-stream)
• oa resume <session_id> (alias of oa chat --resume <session_id>)
• oa logs <session_id> (summarize events.jsonl)

Sessions are stored under ~/.openagentic-sdk by default (override with OPENAGENTIC_SDK_HOME).

Publishing

See docs/publishing.md.

Usage

Streaming:

import asyncio
from openagentic_sdk import OpenAgenticOptions, query
from openagentic_sdk.providers import OpenAIProvider
from openagentic_sdk.permissions import PermissionGate


async def main() -> None:
    options = OpenAgenticOptions(
        provider=OpenAIProvider(),
        model="gpt-4.1-mini",
        api_key="...",  # OpenAI API key
        permission_gate=PermissionGate(permission_mode="prompt", interactive=True),
        setting_sources=["project"],
    )

    async for event in query(prompt="Find TODOs in this repo", options=options):
        print(event.type)


asyncio.run(main())

One-shot:

import asyncio
from openagentic_sdk import OpenAgenticOptions, run
from openagentic_sdk.providers import OpenAIProvider
from openagentic_sdk.permissions import PermissionGate


async def main() -> None:
    options = OpenAgenticOptions(
        provider=OpenAIProvider(),
        model="gpt-4.1-mini",
        api_key="...",
        permission_gate=PermissionGate(permission_mode="callback", approver=lambda *_: True),
    )
    result = await run(prompt="Explain this project", options=options)
    print(result.final_text)


asyncio.run(main())

OpenAI-compatible backend (the examples default to RIGHTCODE):

from openagentic_sdk import OpenAgenticOptions, run
from openagentic_sdk.providers.openai_compatible import OpenAICompatibleProvider
from openagentic_sdk.permissions import PermissionGate

options = OpenAgenticOptions(
    provider=OpenAICompatibleProvider(base_url="https://www.right.codes/codex/v1"),
    model="gpt-5.2",
    api_key="...",  # RIGHTCODE_API_KEY
    cwd=".",
    permission_gate=PermissionGate(permission_mode="prompt", interactive=True),
    setting_sources=["project"],
)

Built-in tools

Default registry includes:

• Read, Write, Edit
• Glob, Grep
• Bash
• WebFetch
• WebSearch (Tavily; requires TAVILY_API_KEY)
• TodoWrite
• SlashCommand (loads .claude/commands/<name>.md)
• Skill (CAS-style single tool for .claude/skills/**/SKILL.md)
• SkillList, SkillLoad, SkillActivate (legacy/compat)

For OpenAI-compatible providers, tool schemas include long-form “how to use this tool” descriptions (opencode-style) to make the model follow rules more reliably.

.claude compatibility

When setting_sources=["project"], the SDK can index:

• CLAUDE.md or .claude/CLAUDE.md (memory)
• .claude/skills/**/SKILL.md
• .claude/commands/*.md

When setting_sources=["project"], query() prepends a system message with project memory + skills/commands index; SkillActivate adds an "Active Skills" section persisted via skill.activated events (survives resume).

Console output (human-first)

Examples use openagentic_sdk.console.ConsoleRenderer, which:

• Prints assistant text by default (human-friendly).
• In debug mode (--debug or OPENAGENTIC_SDK_CONSOLE_DEBUG=1), prints tool/hook/result summaries.

Try the interactive CLI chat example:

• python3 example/45_cli_chat.py

Event compatibility

• events.jsonl is forward-compatible for added fields: deserialization ignores unknown keys on known event types.
• Unknown event types raise openagentic_sdk.errors.UnknownEventTypeError.