Amp Now Lets You Run Coding Agent Sandboxes on Your Own Servers Without Inbound Ports
The Amp coding agent needs its own virtual machine, called an orb, where it clones the repository and runs tests. Now orbs can run on your own servers. The amp orb provider program takes commands from Amp to create, pause, or delete a machine. You do not need to open inbound ports, because the provider connects to ampcode.com on its own. For teams that cannot move code outside their perimeter, a short Docker script is enough.
AI-processed from Amp News (Sourcegraph); edited by Hamidun News
A coding agent does not work in a vacuum. It needs a machine where it can clone a repository, install dependencies, run tests, and pick the task back up an hour or a day later. In Amp, the coding agent from Sourcegraph, these machines are called orbs: one per thread, meaning one per working session. Previously, the service ran them itself. Now the company has opened self-hosted orbs: you supply the compute, and Amp turns it into orbs.
The most important point for security teams: you do not need to open inbound ports in your network. The connection is always initiated by the program inside the perimeter, not by the external service. If your company does not allow code and working files to leave its own infrastructure, this removes the main obstacle to adopting the agent.
What your platform needs
Amp lists four requirements for the compute platform. It must start new machines quickly, pause them and wake them just as quickly, keep them for a long time, and scale to zero. The last point comes down to a practical example: sleeping machines pile up over time, and nobody wants to pay for idle resources. If your platform can do this, you can reproduce the orb experience in-house.
How it works
Inside your infrastructure, you run the amp orb provider command. It opens a connection to ampcode.com on its own and waits for three kinds of requests: spawn (create or resume an orb), pause, and delete. When an event arrives, the provider brings up the orb for the relevant thread. The orb runs in your infrastructure and connects back to its thread on ampcode.com.
You decide how the requests are carried out. The provider is given the path to your program, for example amp orb provider ./dispatch.sh. The script receives the operation and parameters through environment variables: ORB_OPERATION, ORB_ID, ORB_IMAGE, ORB_TOKEN, and, if needed, AMP_URL.
Amp shows several examples for different platforms:
- Docker:
docker runcreates the container,docker stoppauses it, anddocker rm -fdeletes it. At the end, the script runsamp orb bootstrapinside the container. - E2B: first a template with 4 CPU cores and 8 GB of memory is built, then a Node.js script creates the sandbox, pauses it, or deletes it.
- Apple containers: a container starts with 4 cores and 8 GB of memory, and the
start,stop, anddeletecommands map to spawn, pause, and delete. - Modal: the platform has no built-in pause, so on
pausethe script takes a filesystem snapshot and terminates the machine, and on the nextspawnit brings the machine up from that snapshot. - Vercel, Daytona, exe.dev, and Kubernetes: the article also has separate tabs with examples for these.
Image and limitations
The examples use a ready-made image, passed in the ORB_IMAGE variable. It is Amp's own image, and it already contains everything needed to run amp orb bootstrap.
There are caveats too. The control plane still lives on ampcode.com, so this is not a fully autonomous installation. The article gives no prices and no measurements of machine startup speed. The Docker example shows the simplest version: after a pause, the machine's state stays only on the local disk. The Modal example shows a different cost: restoring from a snapshot adds work to pause and spawn.
What to do about it
If you are trying coding agents in your company, check three things. First, can your platform start and pause machines quickly and scale to zero? Second, where will snapshots be stored if there is no built-in pause? Third, will your security team accept a scheme where the connection always originates from inside the network?
For a quick test, the Docker script from the example is enough. It is short, and it shows what the whole task consists of: the machine lifecycle reduces to three operations that you can implement in one small script. For production workloads, choose a platform where pause and resume are already built in, like E2B, or plan snapshots in advance, as in the Modal example.
- Nov 21, 2026OpenAI GPT-6 Cyber Found About 35 Flaws in iPhone, Mac and Other Apple Devices
- Nov 21, 2026Apple's macOS 27.3 beta 2 will let you limit AI agents' access to Mac files
- Nov 21, 2026Apple Is Building an AI Agent for Mac While Making It Harder for Outsiders to Get In: 14 Clicks Instead of 4
- Nov 21, 2026Apple Is Recruiting 200,000 Volunteers to Test a Health AI Chatbot
Key Questions Answered
What your platform needs?
Amp lists four requirements for the compute platform. It must start new machines quickly, pause them and wake them just as quickly, keep them for a long time, and scale to zero. The last point comes down to a practical example: sleeping machines pile up over time, and nobody wants to pay for idle resources. If your platform can do this, you can reproduce the orb experience in-house.
How it works?
Inside your infrastructure, you run the amp orb provider command. It opens a connection to ampcode.com on its own and waits for three kinds of requests: spawn (create or resume an orb), pause, and delete. When an event arrives, the provider brings up the orb for the relevant thread. The orb runs in your infrastructure and connects back to its thread on ampcode.com. You decide how the requests are carried out. The provider is given the path to your program, for example amp orb provider ./dispatch.sh.
What are the key details on Image and limitations?
The examples use a ready-made image, passed in the ORB_IMAGE variable. It is Amp's own image, and it already contains everything needed to run amp orb bootstrap. There are caveats too. The control plane still lives on ampcode.com, so this is not a fully autonomous installation. The article gives no prices and no measurements of machine startup speed. The Docker example shows the simplest version: after a pause, the machine's state stays only on the local disk.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.