CodeArkt is a battery-included implementation of the CodeAct framework with support for the multi-agent architecture. Ship autonomous agents that can reason, write, execute & iterate over code. All from a single Python package.

• Multi-agent orchestration – coordinate hierarchies of specialist agents.
• Secure Python sandbox – secure, ephemeral Docker execution environment for code actions.
• First-class tool ecosystem – auto-discover & register MCP tools.
• Drop-dead simple UI – launch an elegant Gradio chat or run the terminal client.
• Production ready – typed codebase (mypy --strict), CI, tests, Docker & Apache-2.0 license.

Install the package:

pip install codearkt  # requires Python ≥ 3.12

Run your MCP servers:

python -m academia_mcp --port 5056 # just an example MCP server

Run a server with a simple agent and connect it to your MCP servers:

import os
from codearkt.codeact import CodeActAgent
from codearkt.llm import LLM
from codearkt.server import run_server

# Use your own or remote MCP servers
mcp_config = {
    "mcpServers": {"academia": {"url": "http://0.0.0.0:5056/mcp", "transport": "streamable-http"}}
}

# Create an agent definition
api_key = os.getenv("OPENROUTER_API_KEY", "")
assert api_key, "Please provide OpenRouter API key!"
agent = CodeActAgent(
    name="manager",
    description="A simple agent",
    llm=LLM(model_name="deepseek/deepseek-chat-v3-0324", api_key=api_key),
    tool_names=["arxiv_download", "arxiv_search"],
)

# Run the server with MCP proxy and agentic endpoints
run_server(agent, mcp_config, port=5055)

Now run a Python client:

from codearkt.client import query_agent
from codearkt.llm import ChatMessage

history = [ChatMessage(role="user", content="Find an abstract of the 2402.01030 paper")]

for event in query_agent(history, port=5055):
    if event.content:
        print(event.content, end="", flush=True)

Within seconds, you will see agents collaborating, executing Python snippets, and streaming the results back to your console.

You can also use existing clients, Gradio and terminal:

uv run -m codearkt.terminal --port 5055
uv run -m codearkt.gradio --port 5055

For now, explore the well-typed source code.

codearkt/
├─ codeact.py          # Core agent logic
├─ python_executor.py  # Secure sandbox for arbitrary code
├─ event_bus.py        # Pub/Sub for agent events
├─ gradio.py           # Optional web UI
└─ ...
examples/
└─ multi_agent/        # End-to-end usage demos

Pull requests are welcome! Please:

1. Fork the repo & create your branch: git checkout -b feature/my-feature
2. Install dev deps: make install
3. Run the linter & tests: make validate && make test
4. Submit a PR and pass the CI.

Join the discussion in Discussions or open an Issue.

CodeArkt is released under the Apache License 2.0 – see the LICENSE file for details.