Python Sdk

Python SDK for inference.sh - run AI apps, build agents, and integrate with 150+ models. Package: inferencesh (pip install inferencesh). Supports sync/async,...

Security Scan

The SKILL.md is internally consistent with a Python SDK for inference.sh — examples, agents, file uploads, and tool-builder patterns match the description — but there are a few documentation/metadata discrepancies and obvious operational risks (API key use, file uploads, webhook/web requests, and example use of eval) that you should understand before using it.

Purpose & Capability

The name/description (Python SDK for inference.sh) matches the content: examples for sync/async clients, agents, tool-builder, file uploads, sessions, and streaming. The allowed tools (pip, python) and examples align with the declared purpose.

Instruction Scope

The runtime instructions are focused on SDK usage and remain within the SDK's purpose. However, the docs include examples that read local files, auto-upload file paths, use multipart uploads, and show how to enable code execution and eval-style handlers — all legitimate SDK features but they expand the agent's I/O surface and can send local data to an external service if used. The SKILL.md also shows webhook tools that call arbitrary external endpoints.

Install Mechanism

This is instruction-only with no install spec and no code shipped in the skill bundle, so nothing is written to disk by the skill itself. The documented install (pip install inferencesh) is standard for a Python SDK but is external to the skill bundle.

Credentials

The skill metadata declares no required environment variables, but the SKILL.md repeatedly shows using an API key (api_key parameter or INFERENCE_API_KEY env var). That mismatch is a documentation/metadata inconsistency: in practice the SDK needs an inference.sh API key to function. Examples also reference webhook secrets and third-party tokens (e.g., GITHUB_TOKEN, Slack webhooks) as part of examples—expected for integrations but not declared by the skill metadata.

Persistence & Privilege

The skill does not request always:true, no install, no config path access, and does not attempt to modify other skills or system-wide settings. Autonomous invocation is allowed (default) but is not combined with other elevated privileges here.

Assessment

This skill appears to be a documentation-only SDK integration for inference.sh and is coherent with that purpose — but take the following precautions before using it: 1) The SDK requires an inference.sh API key (examples use INFERENCE_API_KEY) even though the skill metadata lists no required env vars; treat that key like any secret and limit its scope. 2) Examples show automatic upload of local file paths and large-file multipart uploads — avoid pointing the SDK at sensitive files unless you intend to send them to the external service. 3) Webhook tools and public=true file options can make data reachable to third parties; verify webhook URLs and permissions. 4) Several examples enable code execution and show use of eval — running untrusted expressions or auto-executing code from agents is dangerous; require human approval and sanitize inputs. 5) The skill bundle is instruction-only and has no provenance (homepage unknown, owner id only). Before installing or running pip install inferencesh, verify the package source (official website or PyPI project page) and review the actual pip package contents and version to ensure you’re installing the legitimate SDK. If you need higher assurance, ask the publisher for a homepage or source repository link and inspect the released package code.

Build AI applications with the inference.sh Python SDK.

Quick Start

pip install inferencesh

from inferencesh import inference

client = inference(api_key="inf_your_key")

# Run an AI app
result = client.run({
    "app": "infsh/flux-schnell",
    "input": {"prompt": "A sunset over mountains"}
})
print(result["output"])

Installation

# Standard installation
pip install inferencesh

# With async support
pip install inferencesh[async]

Requirements: Python 3.8+

Authentication

import os
from inferencesh import inference

# Direct API key
client = inference(api_key="inf_your_key")

# From environment variable (recommended)
client = inference(api_key=os.environ["INFERENCE_API_KEY"])

Get your API key: Settings → API Keys → Create API Key

Running Apps

Basic Execution

result = client.run({
    "app": "infsh/flux-schnell",
    "input": {"prompt": "A cat astronaut"}
})

print(result["status"])  # "completed"
print(result["output"])  # Output data

Fire and Forget

task = client.run({
    "app": "google/veo-3-1-fast",
    "input": {"prompt": "Drone flying over mountains"}
}, wait=False)

print(f"Task ID: {task['id']}")
# Check later with client.get_task(task['id'])

Streaming Progress

for update in client.run({
    "app": "google/veo-3-1-fast",
    "input": {"prompt": "Ocean waves at sunset"}
}, stream=True):
    print(f"Status: {update['status']}")
    if update.get("logs"):
        print(update["logs"][-1])

Run Parameters

Parameter	Type	Description
app	string	App ID (namespace/name@version)
input	dict	Input matching app schema
setup	dict	Hidden setup configuration
infra	string	'cloud' or 'private'
session	string	Session ID for stateful execution
session_timeout	int	Idle timeout (1-3600 seconds)

File Handling

Automatic Upload

result = client.run({
    "app": "image-processor",
    "input": {
        "image": "/path/to/image.png"  # Auto-uploaded
    }
})

Manual Upload

from inferencesh import UploadFileOptions

# Basic upload
file = client.upload_file("/path/to/image.png")

# With options
file = client.upload_file(
    "/path/to/image.png",
    UploadFileOptions(
        filename="custom_name.png",
        content_type="image/png",
        public=True
    )
)

result = client.run({
    "app": "image-processor",
    "input": {"image": file["uri"]}
})

Sessions (Stateful Execution)

Keep workers warm across multiple calls:

# Start new session
result = client.run({
    "app": "my-app",
    "input": {"action": "init"},
    "session": "new",
    "session_timeout": 300  # 5 minutes
})
session_id = result["session_id"]

# Continue in same session
result = client.run({
    "app": "my-app",
    "input": {"action": "process"},
    "session": session_id
})

Agent SDK

Template Agents

Use pre-built agents from your workspace:

agent = client.agent("my-team/support-agent@latest")

# Send message
response = agent.send_message("Hello!")
print(response.text)

# Multi-turn conversation
response = agent.send_message("Tell me more")

# Reset conversation
agent.reset()

# Get chat history
chat = agent.get_chat()

Ad-hoc Agents

Create custom agents programmatically:

from inferencesh import tool, string, number, app_tool

# Define tools
calculator = (
    tool("calculate")
    .describe("Perform a calculation")
    .param("expression", string("Math expression"))
    .build()
)

image_gen = (
    app_tool("generate_image", "infsh/flux-schnell@latest")
    .describe("Generate an image")
    .param("prompt", string("Image description"))
    .build()
)

# Create agent
agent = client.agent({
    "core_app": {"ref": "infsh/claude-sonnet-4@latest"},
    "system_prompt": "You are a helpful assistant.",
    "tools": [calculator, image_gen],
    "temperature": 0.7,
    "max_tokens": 4096
})

response = agent.send_message("What is 25 * 4?")

Available Core Apps

Model	App Reference
Claude Sonnet 4	infsh/claude-sonnet-4@latest
Claude 3.5 Haiku	infsh/claude-haiku-35@latest
GPT-4o	infsh/gpt-4o@latest
GPT-4o Mini	infsh/gpt-4o-mini@latest

Tool Builder API

Parameter Types

from inferencesh import (
    string, number, integer, boolean,
    enum_of, array, obj, optional
)

name = string("User's name")
age = integer("Age in years")
score = number("Score 0-1")
active = boolean("Is active")
priority = enum_of(["low", "medium", "high"], "Priority")
tags = array(string("Tag"), "List of tags")
address = obj({
    "street": string("Street"),
    "city": string("City"),
    "zip": optional(string("ZIP"))
}, "Address")

Client Tools (Run in Your Code)

greet = (
    tool("greet")
    .display("Greet User")
    .describe("Greets a user by name")
    .param("name", string("Name to greet"))
    .require_approval()
    .build()
)

App Tools (Call AI Apps)

generate = (
    app_tool("generate_image", "infsh/flux-schnell@latest")
    .describe("Generate an image from text")
    .param("prompt", string("Image description"))
    .setup({"model": "schnell"})
    .input({"steps": 20})
    .require_approval()
    .build()
)

Agent Tools (Delegate to Sub-agents)

from inferencesh import agent_tool

researcher = (
    agent_tool("research", "my-org/researcher@v1")
    .describe("Research a topic")
    .param("topic", string("Topic to research"))
    .build()
)

Webhook Tools (Call External APIs)

from inferencesh import webhook_tool

notify = (
    webhook_tool("slack", "https://hooks.slack.com/...")
    .describe("Send Slack notification")
    .secret("SLACK_SECRET")
    .param("channel", string("Channel"))
    .param("message", string("Message"))
    .build()
)

Internal Tools (Built-in Capabilities)

from inferencesh import internal_tools

config = (
    internal_tools()
    .plan()
    .memory()
    .web_search(True)
    .code_execution(True)
    .image_generation({
        "enabled": True,
        "app_ref": "infsh/flux@latest"
    })
    .build()
)

agent = client.agent({
    "core_app": {"ref": "infsh/claude-sonnet-4@latest"},
    "internal_tools": config
})

Streaming Agent Responses

def handle_message(msg):
    if msg.get("content"):
        print(msg["content"], end="", flush=True)

def handle_tool(call):
    print(f"\n[Tool: {call.name}]")
    result = execute_tool(call.name, call.args)
    agent.submit_tool_result(call.id, result)

response = agent.send_message(
    "Explain quantum computing",
    on_message=handle_message,
    on_tool_call=handle_tool
)

File Attachments

# From file path
with open("image.png", "rb") as f:
    response = agent.send_message(
        "What's in this image?",
        files=[f.read()]
    )

# From base64
response = agent.send_message(
    "Analyze this",
    files=["data:image/png;base64,iVBORw0KGgo..."]
)

Skills (Reusable Context)

agent = client.agent({
    "core_app": {"ref": "infsh/claude-sonnet-4@latest"},
    "skills": [
        {
            "name": "code-review",
            "description": "Code review guidelines",
            "content": "# Code Review\n\n1. Check security\n2. Check performance..."
        },
        {
            "name": "api-docs",
            "description": "API documentation",
            "url": "https://example.com/skills/api-docs.md"
        }
    ]
})

Async Support

from inferencesh import async_inference
import asyncio

async def main():
    client = async_inference(api_key="inf_...")

    # Async app execution
    result = await client.run({
        "app": "infsh/flux-schnell",
        "input": {"prompt": "A galaxy"}
    })

    # Async agent
    agent = client.agent("my-org/assistant@latest")
    response = await agent.send_message("Hello!")

    # Async streaming
    async for msg in agent.stream_messages():
        print(msg)

asyncio.run(main())

Error Handling

from inferencesh import RequirementsNotMetException

try:
    result = client.run({"app": "my-app", "input": {...}})
except RequirementsNotMetException as e:
    print(f"Missing requirements:")
    for err in e.errors:
        print(f"  - {err['type']}: {err['key']}")
except RuntimeError as e:
    print(f"Error: {e}")

Human Approval Workflows

def handle_tool(call):
    if call.requires_approval:
        # Show to user, get confirmation
        approved = prompt_user(f"Allow {call.name}?")
        if approved:
            result = execute_tool(call.name, call.args)
            agent.submit_tool_result(call.id, result)
        else:
            agent.submit_tool_result(call.id, {"error": "Denied by user"})

response = agent.send_message(
    "Delete all temp files",
    on_tool_call=handle_tool
)

Reference Files

• Agent Patterns - Multi-agent, RAG, human-in-the-loop patterns
• Tool Builder - Complete tool builder API reference
• Streaming - Real-time progress updates and SSE handling
• File Handling - Upload, download, and manage files
• Sessions - Stateful execution with warm workers
• Async Patterns - Parallel processing and async/await

Related Skills

# JavaScript SDK
npx skills add inference-sh/skills@javascript-sdk

# Full platform skill (all 150+ apps via CLI)
npx skills add inference-sh/skills@inference-sh

# LLM models
npx skills add inference-sh/skills@llm-models

# Image generation
npx skills add inference-sh/skills@ai-image-generation

Documentation

• Python SDK Reference - Full API documentation
• Agent SDK Overview - Building agents
• Tool Builder Reference - Creating tools
• Authentication - API key setup
• Streaming - Real-time updates
• File Uploads - File handling