A safe place for your AI agents to run code.
Your agent lives on Edge Compute. It does the risky work in a sandbox.
Starts for one job
The agent creates a sandbox for a task and deletes it when the task is done.
Runs what it needs
Install a library, run a script, open a file nobody on your team wrote.
Separate from your agent
A sandbox has its own memory and files, apart from your agent. It gets only the secrets you pass in.
Runs in isolation
Code runs in its own isolated environment, away from your agent. It can still reach the services you allow.
Most teams add a second vendor for code execution, then wire it to the agent with more credentials. CPU Sandbox is part of the Telnyx Agent Platform: build the agent, deploy it where it can be reached, and give it a place to do work.
Each sandbox is its own microVM with its own kernel. Code your agent wrote, or a file a customer uploaded, runs there and not inside your agent.
Pass secrets in as environment variables, or keep them out and have the real value added at the network edge. You decide which destinations a sandbox can reach.
The agent, its deploy path and its workspace sit in one account. There is nothing extra to buy, secure and connect.
Running a snippet and throwing the machine away is the easy version. Real tasks take many steps, and rebuilding the environment each time loses whatever was in memory.
Isolated runtime
Each sandbox has its own environment, separate from your application and from other customers.
Work that persists
The environment carries across the agent's steps, so tools and dependencies are installed once.
Pause and resume
Checkpoint running work and pick it up later with memory and processes intact, not only the saved files. Paused sandboxes are kept for 7 days by default, subject to configuration.
Fork
Clone a running sandbox into as many as 20 live copies in one request.
Your image
Start from any public linux/amd64 container image. Save templates for repeat use.
Programmatic control
Create a sandbox, run commands, move files in and out and manage its lifecycle through a REST API.
Four API calls cover a task.
# $CONTROL_API = https://api.telnyx.com/v2/compute/sandbox
# 1. Create a sandbox from your image
curl -X POST https://api.telnyx.com/v2/compute/sandbox \
-H "Authorization: Bearer $TELNYX_API_KEY" \
-H "Content-Type: application/json" \
-d '{ "image": { "ref": "docker.io/library/python:3.12-slim" }, "name": "my-task" }'
# 2. Run a command in it (execs go to the sandbox data plane with its token)
curl -X POST $DATA_API/$SANDBOX_ID/exec \
-H "Authorization: Bearer $SANDBOX_TOKEN" \
-H "Content-Type: application/json" \
-d '{ "command": ["python3", "/workspace/task.py"] }'
# 3. Pause it, resume it or fork it
curl -X POST $CONTROL_API/$SANDBOX_ID/pause -H "Authorization: Bearer $TELNYX_API_KEY"
curl -X POST $CONTROL_API/$SANDBOX_ID/connect \
-H "Authorization: Bearer $TELNYX_API_KEY" -H "Content-Type: application/json" \
-d '{ "timeout": 900 }'
curl -X POST $CONTROL_API/$SANDBOX_ID/fork \
-H "Authorization: Bearer $TELNYX_API_KEY" -H "Content-Type: application/json" \
-d '{ "count": 5 }'
# 4. Terminate it when the job is done
curl -X POST $CONTROL_API/$SANDBOX_ID/terminate -H "Authorization: Bearer $TELNYX_API_KEY"CPU Sandbox is one of six products for building, deploying and operating agents.

S3-compatible, AI-ready cloud storage API.
Learn moreBuild agents in code, with the Telnyx harness or your framework.
Learn moreThe runtime your agent lives on.
Learn more
The default models for agents on Telnyx.
Learn more

CPU Sandbox is an isolated, programmable computer an AI agent can start on demand to run code, install tools and work with files. It is part of the Telnyx Agent Platform and is in beta.
Each sandbox is its own microVM with its own kernel. You set which network destinations it can reach, and you choose which secrets, if any, it receives.
Memory, running processes and disk are kept, and the sandbox resumes where it stopped. Paused sandboxes are kept for 7 days by default, subject to configuration.
Any public linux/amd64 container image. Sandboxes are CPU only, with 1 or 2 vCPU and up to 4 GiB of memory in the beta.
Through a REST API. Create a sandbox, run commands, read and write files, pause, fork and terminate.
An agent opens one through the API: a single create call starts a sandbox from your image, execs run the work inside it, and a connect call at the start of each turn resumes a paused sandbox and extends its lifetime. The same flow is available through the telnyx-edge CLI.
They are separate products. CPU Sandbox is a lightweight computer for an agent's code and files. GPU Sandbox is a whole GPU node you rent by the hour for inference workloads.