buildfunctions 0.2.5

The Buildfunctions SDK for Agents: Hardware-isolated CPU and GPU Sandboxes with runtime controls for untrusted AI actions

The Buildfunctions SDK for Agents

 

Hardware-isolated execution environments for AI agents — with runtime controls to help keep unattended runs bounded

Installation

pip install buildfunctions

Quick Start

1. Create an API Token

Get your API token at buildfunctions.com/settings

2. CPU Function

from buildfunctions import Buildfunctions, CPUFunction

client = await Buildfunctions({"apiToken": API_TOKEN})

deployed_function = await CPUFunction.create({
    "name": "my-cpu-function",
    "code": "./cpu_function_code.py",
    "language": "python",
    "memory": 128,
    "timeout": 30,
})

print(f"Endpoint: {deployed_function.endpoint}")

await deployed_function.delete()

3. CPU Sandbox

from buildfunctions import Buildfunctions, CPUSandbox

client = await Buildfunctions({"apiToken": API_TOKEN})

sandbox = await CPUSandbox.create({
    "name": "my-cpu-sandbox",
    "language": "python",
    "code": "/path/to/code/cpu_sandbox_code.py",
    "memory": 128,
    "timeout": 30,
})

result = await sandbox.run()
print(f"Result: {result}")

await sandbox.delete()

4. GPU Function

from buildfunctions import Buildfunctions, GPUFunction

client = await Buildfunctions({"apiToken": API_TOKEN})

deployed_function = await GPUFunction.create({
    "name": "my-gpu-function",
    "code": "/path/to/code/gpu_function_code.py",
    "language": "python",
    "gpu": "T4",
    "vcpus": 30,
    "memory": "50000MB",
    "timeout": 300,
    "requirements": ["transformers==4.47.1", "torch", "accelerate"],
})

print(f"Endpoint: {deployed_function.endpoint}")

await deployed_function.delete()

5. GPU Sandbox with Local Model

from buildfunctions import Buildfunctions, GPUSandbox

client = await Buildfunctions({"apiToken": API_TOKEN})

sandbox = await GPUSandbox.create({
    "name": "my-gpu-sandbox",
    "language": "python",
    "memory": 10000,
    "timeout": 300,
    "vcpus": 6,
    "code": "./gpu_sandbox_code.py",
    "model": "/path/to/models/Qwen/Qwen3-8B",
    "requirements": "torch",
})

result = await sandbox.run()
print(f"Response: {result}")

await sandbox.delete()

Runtime Controls: Help Keep Your Agent Running Unattended

Wrap any tool call with composable guardrails — no API key required, no sandbox needed. RuntimeControls works standalone around your own functions, or combined with Buildfunctions sandboxes.

Available control layers (configure per workflow): retries with backoff, per-run tool-call budgets, circuit breakers, loop detection, timeout + cancellation, policy gates, injection guards, idempotency, concurrency locks, and event-based observability via event sinks.

1. Wrap Any Tool Call (No API Key)

import httpx
from buildfunctions import RuntimeControls

controls = RuntimeControls.create({
    "maxToolCalls": 50,
    "timeoutMs": 30_000,
    "retry": {"maxAttempts": 3, "initialDelayMs": 200, "backoffFactor": 2},
    "loopBreaker": {"warningThreshold": 5, "quarantineThreshold": 8, "stopThreshold": 12},
    "onEvent": lambda event: print(f"[controls] {event['type']}: {event['message']}"),
})

# Wrap any function — an API call, a shell command, an LLM tool invocation
async def run_api(args, runtime):
    payload = args[0]
    async with httpx.AsyncClient() as client:
        response = await client.post("https://api.example.com/data", json=payload)
        return response.json()

guarded_fetch = controls.wrap({
    "toolName": "api-call",
    "runKey": "agent-run-1",
    "destination": "https://api.example.com",
    "run": run_api,
})

result = await guarded_fetch({"query": "latest results"})
print(result)

# Reset budget counters when starting a new run
await controls.reset("agent-run-1")

2. With Hardware-Isolated Sandbox + Agent Safety

import re
from buildfunctions import Buildfunctions, CPUSandbox, RuntimeControls, applyAgentLogicSafety

await Buildfunctions({"apiToken": API_TOKEN})

sandbox = await CPUSandbox.create({
    "name": "guarded-sandbox",
    "language": "python",
    "code": "./my_handler.py",
    "memory": 128,
    "timeout": 30,
})

controls = RuntimeControls.create(
    applyAgentLogicSafety(
        {
            "maxToolCalls": 20,
            "retry": {"maxAttempts": 2, "initialDelayMs": 200, "backoffFactor": 2},
            "onEvent": lambda event: print(f"[controls] {event['type']}: {event['message']}"),
        },
        {
            "injectionGuard": {
                "enabled": True,
                "patterns": [
                    re.compile(r"ignore\s+previous\s+instructions", re.I),
                    re.compile(r"\brm\s+-rf\b", re.I),
                ],
            },
        },
    )
)

async def run_sandbox(runtime):
    _ = runtime
    return await sandbox.run()

result = await controls.run(
    {
        "toolName": "cpu-sandbox-run",
        "runKey": "sandbox-run-1",
        "destination": sandbox.endpoint,
        "action": "execute",
    },
    run_sandbox,
)

print(f"Result: {result}")
await sandbox.delete()

Full runtime controls documentation: https://www.buildfunctions.com/docs/runtime-controls

Runtime controls are provided as best-effort tools to help manage application behavior and resource usage. They do not guarantee prevention of all unintended outcomes. Users are responsible for monitoring their own workloads. See our Terms of Service for full details.

By using this SDK, you agree to the Terms of Service.

The SDK is currently in beta. If you encounter any issues or have specific syntax requirements, please reach out and contact us at [e-mail], and we’ll work to address them.