rebuno 0.4.7

Python SDK for the Rebuno execution runtime

Rebuno Python SDK

Python client library for the Rebuno agent execution runtime.

Installation

pip install rebuno

Quick Start

from rebuno import Client

client = Client(base_url="http://localhost:8080")

execution = await client.create("my-agent", input={"prompt": "hello"})
print(execution.id)

Client defaults to REBUNO_URL and REBUNO_API_KEY env vars when no args are passed.

Building an Agent

from rebuno import Agent, execution

agent = Agent("my-agent")

async def process(prompt: str) -> dict:
    print("running execution", execution.id)
    return {"answer": prompt.upper()}

agent.run(process)

The handler signature is the input schema — process(prompt: str) makes prompt a required field. Pass a pydantic.BaseModel parameter for validation, or input: dict for the raw claim. Use from rebuno import execution to access the current execution's id, session_id, history, etc.

Tools

from rebuno import tool

@tool("web_search")
async def search(query: str, limit: int = 10) -> list[str]:
    return [...]

Tools are plain module-level functions. The wrapper submits an intent to the kernel before running the body. Hand them to your framework as a list:

graph = create_agent(llm, [search, ...])

Building a Runner

from rebuno import Runner, tool

@tool("compute_heavy")
async def heavy(data: str) -> str:
    return process(data)

Runner("compute-1").run()

The runner advertises every @tool it imports, publishes their schemas to the kernel, and services job assignments. @tool(remote=True) lets you declare a stub in agent code for type-checked imperative calls.

Local MCP

The agent connects to the MCP server directly. Install with pip install rebuno[mcp]:

from rebuno import MCPServer

github = MCPServer(
    "github",
    url="https://api.githubcopilot.com/mcp/",
    headers={"Authorization": "Bearer xxx"},
)

# Inside your handler:
graph = create_agent(llm, [..., *github.tools])

Remote tools (incl. remote MCP)

Discover tools that runners host elsewhere — the agent never holds credentials, never opens MCP transport. Schemas come from the kernel directory.

from rebuno import remote

github_tools  = remote.Tools("github")
compute_tools = remote.Tools("compute")

# Inside your handler:
graph = create_agent(llm, [..., *github_tools, *compute_tools])
# or cherry-pick by short name:
graph = create_agent(llm, [..., github_tools["create_pr"], compute_tools["heavy"]])

Each call routes through the kernel to whichever runner advertises that tool ID. Works for any source — @tool Python functions on a runner, MCP servers hosted by a runner, or both.

Documentation

See the full documentation.

License

MIT