A coding agent your team runs on its own machines
A team rollout repeats the single-machine contract: one manifest per project, one root per machine, and a lock that records what ran. The measured server path is one NVIDIA L4. AWS is a place to rent that GPU, not a tested provider, and Foundry Local is a client runtime, not a multi-user server stack.
One root per project, shared manifests
The manifest travels, not the machine. A project commits its outerloop.yaml and AGENT.md, and each developer or server runs the same pair under its own root. The lock records the host, harness, and model revision that root resolved, so two people can compare what they actually ran. A manifest is small enough to review in a pull request.
Do not share a root between projects. A root holds host state, model rows, and evidence for the runs beneath it, and one root per project keeps removal and evidence tied to the work that produced them.
Two machines that run the same lock are running the same pins, which is what a shared default rests on.
- Commit the manifest and AGENT.md, not the root.
- Pin the setup version in the run command. A newer catalog version is a new decision.
- Keep one root per project, so rm -rf removes exactly one project's state.
- Read the evidence before a setup becomes the team default.
Where the team runtime lives
3 shapes fit a team, and one of them is measured. A shared workstation and a single server run the same contract; only the number of users differs. A hosted API is a different contract, because you do not own the runtime.
| Shape | What fits | What we measured |
|---|---|---|
| Shared workstation | One root per project, with the same manifest and harness as a laptop. | The Apple silicon and Linux CPU pages. No shared-user evidence. |
| Team server | One Linux x64 host with an NVIDIA L4, shared manifests, separate roots. | 3 configurations on one L4 with 15 GB of system RAM. |
| Rented GPU (AWS) | A self-managed L4 in your VPC on a g6 instance. | The L4 hardware rows. |
The L4 hardware is measured. The AWS row is a deployment shape.
AWS: a self-managed L4 in your VPC
An AWS g6 instance with an L4 is the closest cloud shape to the measured server row. You manage the instance, the operating system, the runtime, and the root, and oloop writes nothing outside that root. That is how a cloud box can be self-hosted: you own the runtime even though you rent the hardware.
The L4 setup is measured, and AWS is not tested. The evidence came from an NVIDIA L4 with 15 GB of system RAM, not from a g6 instance. The instance type, the driver, the AMI, and the network stay your own variables.
A VPC, a private subnet, or an IAM boundary is a control surface you configure, and it changes nothing about what this runtime is.
- One instance, one region, one root per project.
- The instance's driver and CUDA version must satisfy the catalog entry's constraints.
- The L4 rows are hardware evidence. The cloud bill and the network are yours.
- Keep the project root on a volume you control, so a stopped instance does not take its run receipts with it.
Cost and ownership
We publish no price. The instance, the storage, the network, and the people who run it set the bill, and we do not know your numbers.
On ownership the answer is concrete: one root per project, one removal command, a lock, and an evidence record for each run. A hosted model API moves the runtime cost into a per-request bill and moves the runtime out of your control, which can be cheaper and is a different trade.
Foundry Local is not a server stack
Foundry Local is a client runtime for one user on a Windows, macOS, or Linux device. It can expose an OpenAI-compatible local server, which is useful for an app on that device. It is not a multi-user server stack: there is no tenant model, no shared runtime control, and no evidence layer for a team.
For a team, use the measured Linux server path. The Windows guide covers the native stack and the WSL2 path for a single developer.
Out of scope
This guide does not estimate concurrency or users per GPU. The measured rows are one probe per configuration on one L4, and a team's load is a separate question.
Data flow, evidence artifacts, and the compliance non-claims live in the secure guide, because a security reviewer asks different questions from the person sizing a server.
Install
Install the CLI, then pick the setup for your machine class.
curl -fsSL https://oloop.run/install.sh | shQuestions
Can a team share one root?
No. A shared root mixes the runs of two projects, and removal takes both with it. Share the manifest and AGENT.md instead.
Is AWS qualified?
No. The L4 hardware is measured; AWS g6 was not tested. The instance, driver, AMI, and network are your unmeasured variables.
How many users can one L4 serve?
No. We did not measure concurrency, so the rows on this page do not answer it.
Does this guide cover compliance?
No. It covers operations and ownership. The secure guide covers data flow, evidence, and the non-claims a reviewer will ask about.