klynx 0.0.23

A sophisticated, universal agentic AI coding framework built on LangGraph and LiteLLM.

Klynx Python SDK for Agent Developmemt

Language / 语言: English | 简体中文

klynx is a Python framework for building coding agents with a default think -> act -> feedback loop.
It supports both:

• A ready-to-use default agent runtime
• A composable graph builder API for custom orchestration

Install (PyPI)

pip install -U klynx

Optional browser tooling (needed only if your tools use browser automation):

playwright install chromium

Quick Start

1) Set your model API key

Example (OpenAI):

export OPENAI_API_KEY="sk-..."

2) Create a model and runtime via builder

from klynx import create_builder, setup_model

model = setup_model("gpt-4o")
builder = create_builder(name="quickstart").react()
agent = builder.build(
    working_dir=".",
    model=model,
    max_iterations=20,
    load_project_docs=False,
)

3) Stream a task and read the final answer

for event in agent.invoke(task="Summarize this repository architecture", thread_id="demo"):
    if event.get("type") == "done":
        print(event.get("answer", ""))

API Exports

from klynx import (
    create_agent,
    create_builder,
    KlynxAgent,
    KlynxGraphBuilder,
    ComposableAgentRuntime,
    setup_model,
    list_models,
    set_tavily_api,
    is_tavily_configured,
    run_terminal_agent_stream,
    run_terminal_ask_stream,
)

Usage Tutorial

Tutorial 1: Direct Q&A mode (builder.ask preset)

from klynx import create_builder, setup_model

model = setup_model("gpt-5.3")
builder = create_builder(name="ask-demo").ask()
agent = builder.build(working_dir=".", model=model)

for event in agent.ask("Explain the key modules in this codebase", thread_id="ask-demo"):
    if event.get("type") == "done":
        print(event.get("answer", ""))

Tutorial 2: Build a composable runtime with create_builder

from klynx import create_builder, setup_model

model = setup_model("gpt-4o")

builder = create_builder(name="demo_builder")
builder.react()

runtime = builder.build(
    working_dir=".",
    model=model,
    max_iterations=12,
)

for event in runtime.invoke(task="Find TODOs and propose fixes", thread_id="builder-demo"):
    if event.get("type") == "done":
        print(event.get("answer", ""))

Tutorial 3: Add custom nodes after the default loop

from klynx import create_builder, setup_model

model = setup_model("gpt-4o")

def post_process(runtime, payload):
    return [{"type": "summary", "content": "Post-processing completed."}]

builder = create_builder(name="pipeline")
builder.react()
builder.add_node("post_process", post_process)
builder.add_edge("klynx_loop", "post_process")

runtime = builder.build(working_dir=".", model=model)
for event in runtime.invoke(task="Refactor this module", thread_id="pipeline-demo"):
    print(event)

Tutorial 4: Manage toolsets dynamically

Both the default agent and builder runtime support tool mutation:

runtime.add_tools("group:core")
runtime.add_tools("group:terminal")
runtime.add_tools("group:tui")
runtime.add_tools("group:network_and_extra")
runtime.add_tools("none")

Built-in Tool Groups

Klynx has 6 built-in tool groups:

• system: state_update, run_subtask, parallel_tool_call
• core: read_file, apply_patch, execute_command, list_directory, search_in_files
• terminal: create_terminal, run_in_terminal, read_terminal, wait_terminal_until, read_terminal_since_last, run_and_wait, exec_command, write_stdin, close_exec_session, check_syntax, launch_interactive_session
• tui: open_tui, read_tui, read_tui_diff, read_tui_region, find_text_in_tui, send_keys, send_keys_and_read, wait_tui_until, close_tui, activate_tui_mode
• network_and_extra: web_search, browser_open, browser_view, browser_act, browser_scroll, browser_screenshot, browser_console_logs
• skills: load_skill

Default loading behavior:

• Default agent startup loads group:system and group:core.
• load_skill availability is controlled by skill_injection_mode: preload hides it, hybrid and tool expose it.

Tutorial 5: Permission modes (workspace / global)

agent = create_builder(name="perm").react().build(working_dir=".", model=model)

# default: workspace sandbox
print(agent.get_permission())

# global mode
agent.set_permission("global")

# back to workspace sandbox
agent.set_sandbox(True)

Tutorial 6: Enable web search tools

from klynx import set_tavily_api, is_tavily_configured

set_tavily_api("tvly-...")
print(is_tavily_configured())  # True

If Tavily API key is not configured, web_search is removed from tool groups and JSON schemas.

Tutorial 7: Skill injection modes

agent = create_builder(name="skills").react().build(
    working_dir=".",
    model=model,
    skill_injection_mode="hybrid",  # preload | tool | hybrid
)

• preload: preload SKILL.md from user input hints, hide load_skill.
• tool: disable preload, rely on load_skill.
• hybrid (default): preload first, keep load_skill fallback.

Tutorial 8: Roll back to a checkpoint

You can inspect checkpoint history for a thread and select a rollback target.
Rollback is one-shot by default: the next invoke/ask resumes from the selected checkpoint.

from klynx import create_builder, setup_model

model = setup_model("gpt-4o")
agent = create_builder(name="rollback").react().build(
    working_dir=".",
    model=model,
    load_project_docs=False,
)

thread_id = "rollback-demo"
list(agent.invoke("task A", thread_id=thread_id))
list(agent.invoke("task B", thread_id=thread_id))

history = agent.get_history(thread_id=thread_id, limit=20)
for item in history:
    print(
        item["display_index"],
        item["checkpoint_id"][:12],
        item["iteration"],
        item["action"],
    )

# Roll back to a selected display index (latest first).
agent.rollback(thread_id=thread_id, target_index=1)

# Optional: include tool-managed file restore.
# agent.rollback(thread_id=thread_id, target_index=1, with_files=True)

# Next run resumes from rollback checkpoint.
list(agent.invoke("task C after rollback", thread_id=thread_id))

# Optional: clear a pending rollback if needed.
agent.cancel_rollback(thread_id=thread_id)

Notes:

• Checkpoint rollback restores agent session state, not all external side effects.
• with_files=True restores file edits recorded through mutation tools (apply_patch).
• Shell commands that mutate files outside these tools are out of scope for automatic restore.

Event Model

invoke(...) and ask(...) produce event dictionaries. Common types include:

• token
• reasoning_token
• tool_exec
• tool_result
• warning
• error
• done

A done event usually contains the final answer and token metrics.

Model Setup Notes

setup_model(...) supports both alias and provider/model usage:

setup_model("gpt-4o")
setup_model("openai", "gpt-4o")
setup_model("deepseek", "deepseek-chat")

Use list_models() to inspect available aliases.

Terminal Helpers

For terminal-only usage, you can use helper stream runners:

• run_terminal_agent_stream(...)
• run_terminal_ask_stream(...)

Related Package

If you want a full command-line and TUI experience, install:

pip install -U klynx-cli