Curie: The open-source, self-hostable platform that turns coding agents into co-workers

Who it's for

• For development teams that want AI collaboration on their own infrastructure: Curie is an open-source, self-hostable platform that runs in your own infrastructure and turns coding agents into co-workers you can assign tasks to and accept output from.
• The core scenario is "delegate a GitHub issue and get an implementation back for review": Curie handles the implementation and review of the issue and returns a pull request or merge request that awaits your team's acceptance.
• It also fits Kubernetes operations collaboration: read-only Kubernetes checks run immediately, while configured mutation-type operations request human approval — a good match for handing over low-risk routine work such as inspections and diagnostics.
• For teams that want the approval step kept in human hands: configured mutation operations require fresh human approval, and which actions need prior approval is up to you to configure.
• Teams that already maintain their own coding agent can adopt it with the official curie init --adopt ./my-existing-agent command.
• When it is not a fit: a machine that cannot run Docker, or no way to obtain an Anthropic or OpenRouter model credential — every prerequisite on the official list is required.

Your first attempt in 3 minutes

Prerequisites: a machine with Docker already running, kind, kubectl, and Helm installed, plus an Anthropic or OpenRouter model credential. If any one of these is missing, you cannot get started following the official instructions.

Run the official installation script in a terminal:

curl -fsSL https://raw.githubusercontent.com/curie-eng/curie/main/get-curie.sh | bash

The script comes from the main branch of the Curie repository; a piped command executes a remote script directly with your user privileges, so if that bothers you, download it first and review its contents before running.

For the first attempt, take the low-risk path: pick a repository issue with a clear description and a small blast radius and delegate it to Curie for implementation and review. When it finishes, it returns a pull request or merge request that awaits your team's acceptance. The success signal is clear: on your code hosting platform you receive the PR/MR it produced and can go through the changes one by one.

How to enter model credentials after installation, or how to link your code hosting account, is not covered step by step in the official instructions; go by the script's output prompts and the repository documentation.

Use cases

Case 1: turn an issue into a PR awaiting review. Scenario: a clearly scoped implementation issue has piled up in the team repository with no one free to handle it in time. How to start: delegate the issue to Curie, which corresponds to the official "delegate the issue for implementation and review" flow. Process and result: when it finishes it hands back a PR/MR, and you review the changes one by one on your own hosting platform; the confirmable endpoint is obtaining a PR/MR that awaits your team's acceptance.

Case 2: read-only Kubernetes inspection. Scenario: you need to know the current state of the cluster but do not want to go through a full change-approval cycle for one routine status check. How to start: ask Curie for a read-only Kubernetes check, which uses exactly the read-only Kubernetes tools described in the official notes. Process and result: read-only operations run immediately and need no extra approval; if you then request a configured mutation-type operation, it first asks for a fresh human approval and only then proceeds. Keep your first inspection within read-only scope and you will not run into an approval wait; to observe the approval step, add one mutation-type request and you will see the approval prompt. Which read-only tools are available and how to invoke them is not itemized in the official notes.

Limits of use

• Environmental prerequisites: Docker running, kind, kubectl, and Helm in place, plus an Anthropic or OpenRouter model credential; OS and hardware architecture requirements are not stated by the official source.
• Side effects and confirmation points: read-only Kubernetes tools run immediately; configured mutation operations request a fresh human approval every time; the output of a delegated implementation is a PR/MR awaiting team acceptance. What the installation script creates on your machine and which directories it occupies are not itemized officially, so you can review the script's contents before executing it.
• Where it runs: a self-hosted platform that runs in your own infrastructure.
• Adopting an existing agent: the official curie init --adopt ./my-existing-agent command is provided; the official documentation does not specify what the command preserves of the existing agent. Initialization and configuration beyond this are not covered by the official instructions.
• What to verify in the repository before acting: the credential configuration steps after installation, how to connect your code hosting account, the complete list of available read-only tools and configurable mutation operations, whether model providers beyond these two are supported, and the exact list of coding agents it can work with. Check each item against the repository documentation before you start.