parallel-google-adk

Google Agent Development Kit (ADK) tools and plugin for Parallel — grounded web search, clean extraction, and cited deep research with structured output.

Install

uv add parallel-google-adk           # or: pip install parallel-google-adk
export PARALLEL_API_KEY=your-key-here # get one at https://platform.parallel.ai

Quickstart

import os
from google.adk.agents import LlmAgent
from google.adk.runners import Runner

from parallel_google_adk import (
    web_search, web_fetch, extract,
    deep_research, enrich,
    ParallelTracingPlugin,
)

agent = LlmAgent(
    model="gemini-flash-latest",
    name="research_agent",
    instruction=(
        "You are a research assistant. Use web_search to find sources, "
        "web_fetch or extract for specific URLs, and deep_research for "
        "comprehensive cited reports."
    ),
    tools=[web_search, web_fetch, extract, deep_research, enrich],
)

# Plugins attach to the Runner, not the agent.
runner = Runner(agent=agent, app_name="my-app", plugins=[ParallelTracingPlugin()])

Tools

Tool	What it does
web_search	Grounded web search returning citation-aware excerpts.
web_fetch	Clean content from a single URL. Handles JS and PDFs.
extract	Batch extraction across 1–20 URLs with optional objective.
deep_research	Multi-hop investigation with per-claim citations. Returns a markdown report.
enrich	Structured (JSON-Schema-conforming) enrichment of a list of entities.

Plugin

ParallelTracingPlugin records latency, citation count, and (when reported) cost for every Parallel tool call. Read the trace from tool_context.state["_parallel_calls"].

MCP alternative

If you only want Search and don't want a Python dependency, the Parallel Search MCP works with ADK's MCPToolset:

import os
from google.adk.tools.mcp_tool.mcp_toolset import MCPToolset
from google.adk.tools.mcp_tool.mcp_session_manager import StreamableHTTPConnectionParams

MCPToolset(
    connection_params=StreamableHTTPConnectionParams(
        url="https://search.parallel.ai/mcp",
        headers={"Authorization": f"Bearer {os.environ['PARALLEL_API_KEY']}"},
    ),
)

This package is the typed-FunctionTool path for users who want fine-grained schemas, hidden polling, structured enrichment, and tracing.

Examples

See examples/research_agent.py for a runnable demo. A run looks like this:

Development

pip install -e ".[dev]"
pytest                              # unit tests (no network)
pytest tests/integration/           # live API smoke (needs PARALLEL_API_KEY + GOOGLE_API_KEY)

License

MIT