mycode-sdk 0.5.2

Multi-turn tool-calling agent runtime for embedding the mycode agent loop.

mycode-sdk

Lightweight Python SDK for building agents.

Install

uv add mycode-sdk
# or
pip install mycode-sdk

Quick start

import asyncio

from mycode import Agent, bash_tool, read_tool


async def main() -> None:
    agent = Agent(
        model="claude-sonnet-4-6",
        api_key="YOUR_API_KEY",
        tools=[read_tool, bash_tool],
    )

    async for event in agent.achat("Read pyproject.toml and tell me the project name."):
        if event.type == "text":
            print(event.data["delta"], end="", flush=True)


asyncio.run(main())

Agent(...) infers the provider from the model id and defaults cwd to the current working directory. No tools are registered unless you pass tools=[...], and nothing is persisted unless you pass session_dir=.

For a synchronous call, use run() — it collects the stream into a RunResult:

result = agent.run("Read pyproject.toml and tell me the project name.")
print(result.text)

Call achat or run again on the same Agent to continue the conversation — history accumulates in agent.messages.

To persist across processes, pass session_dir as the root directory; each session_id becomes a subdirectory. Reconstruct with the same (session_dir, session_id) to resume.

Built-in tools

from mycode import read_tool, write_tool, edit_tool, bash_tool

Only bash_tool streams incremental output as tool_output events; the others return a single result.

Custom tools

@tool wraps a sync or async def Python function as a ToolSpec. Parameter type hints become the JSON schema sent to the provider.

Annotate the first parameter as ToolContext to have the context injected. Use ctx.read / ctx.write / ctx.edit / ctx.bash to invoke the built-ins, or ctx.call(name, args) for any registered tool by name.

from mycode import Agent, ToolContext, read_tool, tool


@tool
def summarize_file(ctx: ToolContext, path: str) -> str:
    """Return the first line of a text file."""

    result = ctx.read(path)
    return result.output.splitlines()[0] if result.output else ""


agent = Agent(
    model="claude-sonnet-4-6",
    api_key="YOUR_API_KEY",
    tools=[read_tool, summarize_file],
)

A bare str return becomes the tool output; any other JSON-serializable value is dumped to JSON. For finer control, return a ToolExecutionResult to set output, content (multimodal blocks such as images), metadata (structured UI data), and is_error independently.

See docs/sdk.md for the event stream, cancellation, session rules, and the full Agent / @tool reference.