blade
Get your API key Check a coding task.
Run a proposed change against your project’s checks in a dedicated environment. Collect the diff and test output for review.
Serverless / Agent sandboxes
Let your agents
run code.
Give AI agents an isolated execution environment for Python code, data analysis and GPU tasks. Run the work, collect the result and bring it back into your application.
Join the Serverless waitlistAccess is opening gradually. Runtime and GPU availability are confirmed during onboarding.
task → environment → resultConceptual workflow
Your agent
“Run this analysis on the sample.”
analyze.pyIsolated task environment
Inputs + dependencies
load → execute → collect GPU when the workload needs it
stdoutoutput filesexit status
Code execution
for AI agents.
Bring a concrete workload to onboarding. Start with a representative input and a clear definition of success.
Turn data into an answer.
Prepare a Python job that cleans data, computes a summary or creates an artifact. Keep inputs and outputs explicit.
Add a GPU step.
Bring workloads that need GPU compute, such as model inference or data processing. Size the runtime and memory with the team.
Prepare the workload
Start with one task.
Make it testable.
Keep the task separate from the agent that requested it. Define the input, pin its dependencies and decide what a successful output looks like.
- Write a small Python entry point with a representative input.
- List required packages, files and any GPU memory needs.
- Share expected duration, traffic and output requirements during onboarding.
analyze.py · local workload examplePython
def summarize(rows):
values = [float(row["amount"]) for row in rows]
return {
"count": len(values),
"total": round(sum(values), 2),
}
# A small, representative input for local testing.
print(summarize([
{"amount": "12.50"},
{"amount": "7.50"},
]))
# Expected: {"count": 2, "total": 20.0}This function runs locally with Python. It illustrates an input/output contract, not the Serverless SDK or a sandbox security boundary. Deployment setup is provided during onboarding.
Define the boundaries
around the code.
Agree these settings with the team before connecting agent-generated work to production.
What does it need?
Python runtime, pinned packages, input files and model weights. Specify GPU memory if the job needs it.
What may it reach?
Required network destinations, data and secrets. Decide how each task gets only the access it needs.
When should it stop?
Timeout, memory budget and concurrency. Define retry behaviour so a repeated job does not duplicate work.
What should remain?
Logs, files and exit status. Confirm how to retrieve them, how long to keep them and when to clean up.
A different starting point / Open Models
Need a model
to do the maths?
For a model-led calculation, Open Models has a built-in code interpreter. Declare the tool and let Blade manage its execution loop.
Requires code interpreter access on your account and a model with native tool calling. This is a managed tool, separate from deploying your own Serverless environment.
Read the tool-calling referencecalculate.pyPython
import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.models.blade.sh/v1",
api_key=os.environ["BLADE_API_KEY"],
)
reply = client.chat.completions.create(
model=os.environ["BLADE_MODEL"],
messages=[{
"role": "user",
"content": "Use Python to find the median of 10, 12, 98.",
}],
extra_body={
"tools": [{"type": "code_interpreter"}],
},
)
print(reply.choices[0].message.content)Install openai and set BLADE_API_KEY and BLADE_MODEL. Select a tool-capable model with access to code execution.
Before you connect an agent.
Can I create a Serverless sandbox today?
Serverless access is opening gradually. Join the waitlist with your workload. Runtime setup, available GPU classes and deployment controls are provided during onboarding.
Is a sandbox the same as a coding model?
No. A model proposes code or a tool call. The execution environment runs the allowed task. Your application or agent orchestrates the steps and reviews the result.
Will my existing agent framework work?
Bring the framework, dependency list and commands you want to run. Compatibility and required permissions are checked during onboarding; this page does not imply a preconfigured Claude, Codex or Hermes integration.
How are sandboxes priced?
GPU classes, scaling settings and commercial terms are confirmed with your Serverless access. Open Models token rates and dynamic pricing are not a quote for a custom sandbox workload.