toolbroker-claude-agent

Claude Agent SDK adapter for ToolBroker.

pip install toolbroker-claude-agent

from claude_agent_sdk import ClaudeAgentOptions, ClaudeSDKClient
from toolbroker import ToolBroker
from toolbroker_claude_agent import ClaudeAgentAdapter

broker = ToolBroker()
broker.add_functions([search_orders, issue_refund, check_inventory])
broker.index()

adapter = ClaudeAgentAdapter()
selection = broker.select("customer wants a refund", k=3)

options = ClaudeAgentOptions(
    mcp_servers={"tools": adapter.create_server(selection.tools)},
    allowed_tools=adapter.allowed_tool_names(selection.tools, server="tools"),
)

render() returns SdkMcpTool objects; create_server() wraps them into the in-process MCP server config the SDK expects.

Why allowed_tool_names exists

The SDK gates tools by name in the form mcp__<server>__<tool>. Getting that string wrong means the model sees the tool and is then refused permission to call it — a confusing failure. The helper builds the names from the same selection, so the two cannot drift apart.

Result shape

Handlers must return MCP content blocks. Anything your function returns is wrapped automatically: a dict that already has content passes through untouched, and everything else becomes a single text block. Override with result_formatter= if you need image or resource blocks.