Developing Genkit Python

GoogleOfficialPublished by Google

Skill
Tool integration
Data and AI
Apache-2.0
Medium risk

From the author

Develop AI-powered applications using Genkit in Python. Use when the user asks about Genkit, AI agents, flows, or tools in Python, or when encountering Genkit errors, import issues, or API problems.

— the description by Google

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

The Genkit Python SDK provides a unified framework for developing AI features, including generation, streaming, tool integration, and multi-turn agents. It emphasizes a developer-centric workflow that integrates with the Genkit CLI for tracing, debugging, and monitoring model interactions. The framework supports advanced agent capabilities such as session management, human-in-the-loop interactions, branching, and background task execution.

What it does

Provides a unified SDK for building AI agents, flows, and generative features in Python.
Supports multi-turn agent interactions with state management and session persistence.
Enables observability through the Genkit CLI for tracing model inputs, outputs, and tool calls.
Facilitates integration with FastAPI for serving agents and flows.
Supports advanced agent patterns like background task execution, branching, and human-in-the-loop approval.
Includes tools for evaluating AI performance and managing prompts.

When to use it

When building AI-powered applications or chatbots in Python.
When you need to implement multi-turn agents with history and state.
When you need to debug AI model interactions, tool calls, or latency issues.
When you require a structured way to define and run generative flows.
When you need to integrate human-in-the-loop or background processing into AI workflows.

Advantages

Unified SDK for diverse AI tasks including generation, agents, and flows.
Powerful CLI-based tracing and debugging that works without modifying application code.
Native support for complex agent patterns like branching and background work.
Flexible session management for persistent agent interactions.
Seamless integration with existing Python web frameworks like FastAPI.

Usage examples

“How do I generate a simple response using Genkit in Python?”: The agent provides a code snippet using the Genkit class and GoogleAI plugin to perform a basic generate call.
“How can I debug my Genkit application to see tool calls?”: The agent explains how to use 'genkit start -- uv run src/main.py' to capture traces and inspect them via the Dev UI or CLI.
“How do I run a specific flow from the command line?”: The agent provides the 'genkit flow:run' command syntax to execute a named flow with specific input data.

Install

Download the skill as a ZIP:
https://mp-staging.autogineer.com/agent-skills/google/developing-genkit-python/download

Unzip it into the cross-client skills folder of your project:

mkdir -p .agents/skills && unzip developing-genkit-python.zip -d .agents/skills

Documentation by Google

Genkit Python

Build AI features in Python — generate, stream, tools, flows, and multi-turn agents — with one SDK.

Prerequisites

Python 3.10+ and uv (install)
Genkit CLI: npm install -g genkit-cli if genkit --version is missing

New app? Setup. Patterns? Examples.

Hello World

from genkit import Genkit
from genkit_google_genai import GoogleAI

ai = Genkit(
    plugins=[GoogleAI()],
    model='googleai/gemini-flash-latest',
)

async def main():
    response = await ai.generate(prompt='Tell me a joke about Python.')
    print(response.text)

if __name__ == '__main__':
    ai.run_main(main())

Agents (Beta)

Multi-turn chats with history, typed state, human approval, branching, and background work. Start here: Agents.

chat = agent.chat()
res = await chat.send('Hello')           # AgentResponse
turn = chat.send_stream('Hello')         # AgentTurn — .stream / .response

More: sessions · HITL · branching · background · state · artifacts · custom · HTTP

Imports

Google AI: from genkit_google_genai import GoogleAI
Agents: from genkit.agent import InMemorySessionStore, ...
Middleware: from genkit_middleware import Middleware, ToolApproval, ...
FastAPI: from genkit_fastapi import serve_agent, serve_flow
Evals: from genkit_evaluators import register_genkit_evaluators

Workflow

Agent or flow? If the task is conversational, multi-turn, or described as "an agent", "assistant", or "chatbot", build it with ai.define_agent (see Agents) rather than hand-rolling a generate + tools loop inside a flow. Reach for a plain flow only for single-shot, stateless generation.
Set GEMINI_API_KEY. Use prefixed model ids (googleai/gemini-flash-latest).
Enter via ai.run_main(main()) for Genkit apps (especially under genkit start). See Common Errors.
Run with Dev Workflow (genkit start + Dev UI).
Verify with traces, not a blind run. Running the app directly (uv run) does not capture dev traces. See Genkit CLI for how to run your app and capture traces.
Stuck? Common Errors first.

Genkit CLI (recommended)

genkit start unintrusively wraps any Python program that uses the Genkit library, running it unchanged while capturing traces from every Genkit action so you can prove tools were actually called and inspect model I/O from the terminal, even for headless checks. It forwards stdio, so interactive CLI tools that rely on stdin/stdout work without issues. Running the app directly (uv run) skips trace capture, so you're debugging blind.

Primary pattern (default): prefix genkit start -- to your normal run command. This collects telemetry from any Genkit code your program runs, whether triggered from the dev UI, your own web server/web UI, or a plain script:

genkit start -- uv run src/main.py
genkit start --noui -- uv run src/main.py   # same, without the Dev UI (still a persistent server)

genkit start runs until you stop it with Ctrl+C. That is expected and correct for the common cases: a server your web/mobile app calls, or an interactive CLI you exit yourself. --noui only drops the Dev UI; it is not a one-shot command and will not exit on its own. Do not use genkit start as a blocking step in automated/non-interactive contexts; use flow:run (below) for that.

Non-interactive use (agents/CI): add the global --non-interactive flag before -- so the CLI uses defaults and never blocks on a prompt (e.g. the first-run analytics notice): genkit start --non-interactive -- uv run src/main.py (works with flow:run too).

Run a flow (flow:run): invoke a specific flow by name from the CLI. Append your run command after -- to spin up the runtime just for this run (the command runs as-is to register your flows):

genkit flow:run myFlow '{"data": "input"}' -- uv run src/main.py

This is self-terminating: it runs the flow once, prints a Trace ID, then exits, so it's the right choice for a quick, non-interactive check (unlike genkit start). Note: flow:run runs flows (@ai.flow()), not agents; you can't flow:run an agent (ai.define_agent) directly. To exercise an agent from the CLI, wrap one turn in a throwaway flow and run that (see Agents).

Debugging with traces: the fastest way to see prompts, model inputs/outputs, tool calls, latencies, and errors. Inspect from the terminal after any run under genkit start:

genkit trace:list                        # find recent trace IDs
genkit trace:get <traceId>               # full trace details (inputs, outputs, tool calls, errors)
genkit trace:get <traceId> --format json # machine-readable JSON, safe to pipe into jq or other parsers

For machine-readable output, pass --format json to get clean JSON you can pipe into jq or other parsers. The default output is human-oriented (banner/log lines, possible truncation on large traces), so don't pipe that form directly; use --format json, grep, or the Dev UI trace viewer.

See Dev Workflow for the full checklist and Dev UI walkthrough.

References

Examples: Structured output, streaming, flows, tools, embeddings.
Setup: New project bootstrap and plugins.
Common Errors: Read first when something breaks.
FastAPI: HTTP, genkit_fastapi_handler, parallel flows.
Dotprompt: .prompt files and helpers.
Evals: Evaluators and datasets.
Dev Workflow: genkit start, Dev UI, checklist.
Agents (Beta): Multi-turn API.

Files

ATTRIBUTION.mdAdded by AutoGineer866 B
LICENSELicence file added by AutoGineer11.1 KiB
SKILL.md6.5 KiB
references/agents-artifacts.md1.5 KiB
references/agents-background.md2.1 KiB
references/agents-branching.md1.3 KiB
references/agents-custom.md3.8 KiB
references/agents-http.md2.9 KiB
references/agents-human-in-the-loop.md3.6 KiB
references/agents-sessions.md1.7 KiB
references/agents-state.md3.7 KiB
references/agents.md5.2 KiB
references/common-errors.md2.6 KiB
references/dev-workflow.md5.5 KiB
references/dotprompt.md4.7 KiB
references/evals.md3.9 KiB
references/examples.md5.3 KiB
references/fastapi.md9.3 KiB
references/setup.md1.9 KiB

references/agents-human-in-the-loop.md

Agent Human-in-the-Loop / Interrupts (Beta)

Beta / preview API. See agents.md.

An interrupt pauses mid-turn so your code (or a human) can decide, then resume. Works with or without a store.

Flow: send → finish_reason == INTERRUPTED + res.interrupts → human input → resume.

ToolApproval

Tools in allowed_tools run freely; everything else pauses. Empty list ⇒ every tool needs approval. When you also mount Filesystem / Artifacts, remember to allow-list (or deliberately interrupt on) write_file / edit_file / write_artifact — the coding-agent example often auto-approves reads only.

from uuid import uuid4

from pydantic import BaseModel, Field
from genkit_google_genai import GoogleAI
from genkit_middleware import Middleware, ToolApproval

from genkit import Genkit, ToolRequestPart
from genkit.agent import AgentFinishReason, InMemorySessionStore

ai = Genkit(plugins=[GoogleAI(), Middleware()])
tool_approval = ToolApproval(allowed_tools=[])


class TransferInput(BaseModel):
    amount: float
    to_account: str = Field(alias='toAccount')


class TransferOutput(BaseModel):
    success: bool
    transaction_id: str = Field(alias='transactionId')


@ai.tool(name='transferMoney', description='Transfer money between accounts.')
async def transfer_money(input: TransferInput) -> TransferOutput:
    return TransferOutput(success=True, transactionId=f'txn-{uuid4().hex[:12]}')


agent = ai.define_agent(
    name='bankingAgent',
    model='googleai/gemini-flash-latest',
    system='Banking assistant. Call transferMoney when the user asks to transfer money.',
    tools=[transfer_money],
    use=[tool_approval],
    store=InMemorySessionStore(),
)

Detect and resume

chat = agent.chat()

out1 = await chat.send('Transfer $500 to account 12345 for rent.')
# finish_reason == INTERRUPTED; transferMoney has NOT executed yet.

restart_parts: list[ToolRequestPart] = [
    intr.restart(resumed_metadata={'tool_approved': True}) for intr in out1.interrupts
]
out2 = await chat.resume(restart=restart_parts)

Resume builders (return parts; you still call chat.resume):

interrupt.restart(...) — re-issue the tool request (ToolApproval after OK). resumed_metadata={'tool_approved': True} (camelCase toolApproved also ok)
interrupt.respond(output) — supply a tool response without running the tool (use this to deny: e.g. respond({'ok': False, 'error': 'denied by policy'}))

# Approve (runs the tool):
await chat.resume(
    restart=[intr.restart(resumed_metadata={'tool_approved': True}) for intr in out1.interrupts]
)

# Deny (model sees the respond payload; tool does not execute):
await chat.resume(
    respond=[intr.respond({'ok': False, 'error': 'export denied'}) for intr in out1.interrupts]
)

await chat.resume(
    respond=[a.respond({'approved': True})],
    restart=[b.restart(resumed_metadata={'tool_approved': True})],
)

turn = chat.resume_stream(restart=restart_parts)
async for chunk in turn.stream:
    if chunk.text:
        print(chunk.text, end='', flush=True)
final = await turn.response

Without a store: keep the same chat, or round-trip messages/state/artifacts (client-managed state).

Build resume parts from res.interrupts only. Loop until finish reason is no longer INTERRUPTED. Don't treat an interrupted turn as the final reply.

For approval UIs, each interrupt exposes .name, .ref, and .input (the pending tool call). Use those to render the prompt; then .restart(...) or .respond(...).

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Needs an API key, Needs Python, Needs Node.js

Security

What our scanner noticed· Medium risk

Instructions that use the network
Downloads piped into a shell

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Published by

Google

Publisher

Google (Official)

Source

google/skills · skills/cloud/developing-genkit-python · commit 2194d3b

Licence

Apache License 2.0 (redistribution allowed)

LICENSE (added by AutoGineer from LICENSE)
ATTRIBUTION.md (written by AutoGineer)

Synced

2026-11-08

AutoGineer is not affiliated with or endorsed by Google.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download developing-genkit-python.zip