Coder: Manage cloud development environments and AI agents with Terraform
Coder gives teams Terraform-managed cloud workspaces and AI agents while keeping model keys out of workspaces.
GitHub coder/coder Updated 2026-09-18 Branch main Stars 14.8K Forks 1.5K
Go TypeScript Terraform Kubernetes AI coding agents Self-hosted

🧭 Decision Guide

Try it if you

  • You need Terraform to standardize EC2, Kubernetes Pod, or Docker development workspaces.
    The README highlights “Define cloud development environments in Terraform” and lists EC2 VMs, Kubernetes Pods, and Docker Containers.
  • You want Claude Code, Codex, or OpenCode to execute coding work on your infrastructure.
    Integrations > Official lists Coding Agents including Claude Code, Codex, and OpenCode; the README says Coder Agents runs on self-hosted infrastructure.
  • You need centralized governance for Anthropic, OpenAI, Google, Bedrock, or self-hosted models.
    The README lists these model sources and states centralized model governance, cost tracking, and audit logging.
  • You need to connect to remote development workspaces through VS Code or JetBrains Toolbox.
    Integrations > Official lists the VS Code Extension and JetBrains Toolbox Plugin.

Skip it if you

  • Your production deployment cannot provide PostgreSQL 13 or later and an external access URL.
    Quickstart explicitly requires “a PostgreSQL database (version 13 or later)” and “an external access URL” for production.
  • Your deployment environment does not include EC2, Kubernetes, Docker, or another workspace infrastructure named by the README.
    The README lists EC2 VMs, Kubernetes Pods, and Docker Containers as workspace infrastructure examples.
  • Your compliance policy cannot accept the GNU Affero General Public License v3.0.
    Project metadata lists the license as GNU Affero General Public License v3.0.

Requirements

  • Linux or macOS can use the install script from the README; Windows can use a binary or installer from GitHub Releases.
  • For evaluation, Coder can use a built-in database and a *.try.coder.app access URL.
  • Production deployment requires PostgreSQL 13 or later and an external access URL.
  • Workspace networking uses a secure WireGuard tunnel.
  • Supported model sources include Anthropic, OpenAI, Google, Bedrock, and self-hosted models.

First step (verbatim from README)

curl -L https://coder.com/install.sh | sh

Watch out

  • Without the production flags, Coder uses a built-in database and sets up a *.try.coder.app access URL.
    Quickstart states that without these flags, Coder uses a built-in database and sets up a *.try.coder.app access URL.
  • The production parameters --postgres-url and --access-url must follow the README example.
    Quickstart gives the production command coder server --postgres-url <url> --access-url <url>.
  • The AI agent loop runs in the control plane rather than inside the workspace.
    The README says the Coder Agents loop executes in the control plane on your infrastructure.
  • LLM credentials stay out of workspaces, while model governance uses components such as AI Gateway.
    The README states no API keys in workspaces and documents AI Gateway for authentication, auditing, and cost controls.

Not stated in the README

  • The README does not provide a supported-version matrix for EC2, Kubernetes, and Docker.
  • The README does not specify minimum CPU, memory, or storage requirements for the control plane, PostgreSQL, or workspaces.
  • The README does not describe WireGuard limitations with proxies, firewalls, or offline environments.
  • The README does not explain functional differences among Anthropic, OpenAI, Google, Bedrock, and self-hosted models.
  • The README does not list the specific Premium features, pricing, or scale limits.
  • The README does not describe the maintenance responsibilities or enterprise-support response times for the 10 contributors.
  • The README does not state how many of the 1,030 issues are open or their severity.

💡 Deep Analysis

6
Yes I have one Linux or macOS machine and want to validate Coder by creating a first workspace from a Docker template. Is the README Quickstart sufficient for an evaluation?
For: A small engineering lead who wants to validate remote development and AI-agent workflows by creating a Docker workspace on one Linux or macOS machine

Yes for evaluation. The README provides a short path covering installation on Linux or macOS, starting the server, creating a Docker template, and provisioning the first workspace. It should not be treated as a production design.

  • The official install.sh installation method is documented for Linux and macOS.
  • After running coder server, the user can open http://localhost:3000 to create the initial user, a Docker template, and the first workspace.
  • Without PostgreSQL and an external access URL, Coder uses a built-in database and configures a *.try.coder.app evaluation URL.
  • The README explicitly positions this configuration for evaluation, while production requires PostgreSQL 13+ and an external access URL.

The Quickstart is therefore sufficient to validate Docker workspaces and basic access. Long-term team or public use needs the production deployment path.

  • Quickstart: "Try Coder with the install script on Linux and macOS"
  • Quickstart: "coder server"
  • Quickstart: "open http://localhost:3000 to create your initial user, create a Docker template, and provision your first workspace"
  • Quickstart: "Without these flags, Coder uses a built-in database and sets up a `*.try.coder.app` access URL for evaluation"
curl -L https://coder.com/install.sh | sh
Not stated in the README:The README does not specify the minimum CPU, memory, or disk requirements for a single-machine Docker evaluation.;The README does not state whether the Quickstart fully validates Coder Agents, VS Code, or JetBrains integrations.
Yes We want to run Claude Code, Codex, or OpenCode on our own infrastructure while forbidding model API keys in workspaces. Do Coder Agents provide identity tracking, model governance, and cost auditing?
For: An AI platform engineer who needs to run Claude Code, Codex, or OpenCode on company infrastructure with centralized model auditing, per-user attribution, and no LLM keys stored in workspaces

Yes. Coder Agents execute their agent loop on the self-hosted control plane and explicitly avoid placing API keys in workspaces. However, the README does not fully define approval workflows for high-risk commands or the compliance boundary of each model service.

  • The README lists Claude Code, Codex, and OpenCode as agents that can run isolated in Coder workspaces.
  • Coder supports Anthropic, OpenAI, Google, Bedrock, and self-hosted models rather than forcing one provider.
  • Its description promises “no LLM credentials in workspaces,” “user identity on every action,” centralized model governance, cost tracking, and audit logging.
  • The AI Gateway section describes centralized authentication, auditing, and cost controls for AI tooling.

Coder is a strong fit when the primary requirement is centralized credentials and attribution. Fine-grained approval, data redaction, and model-specific compliance still require verification.

  • Self-Hosted Cloud Development Environments and AI Agents: "Delegate coding work to AI agents on your infrastructure"
  • Self-Hosted Cloud Development Environments and AI Agents: "Bring any model (Anthropic, OpenAI, Google, Bedrock, self-hosted)"
  • Self-Hosted Cloud Development Environments and AI Agents: "No LLM credentials in workspaces, user identity on every action"
  • Documentation: "AI Gateway: Centralize authentication, auditing, and cost controls for AI tooling"
  • Integrations / Official: "Run agents like Claude Code, Codex, and OpenCode isolated in Coder workspaces"
Not stated in the README:The README does not specify approval granularity for agent commands, network isolation capabilities, or controls for data exfiltration.;The README does not explain how billing data is collected for each model provider or what hardware and deployment requirements apply to self-hosted models.
Yes We already manage EC2 and Kubernetes resources with Terraform and need to support Docker workspaces as well. Is Coder suitable as the unified control plane for multiple development teams?
For: A platform engineer who manages EC2, Kubernetes Pod, and Docker workspaces with Terraform and provides a unified remote environment for multiple development teams

Yes, because Coder combines workspace infrastructure, lifecycle management, and access into one self-hosted platform. The platform team must still operate the control plane and the underlying compute resources.

  • The README states that workspaces are defined with Terraform and explicitly lists EC2 VMs, Kubernetes Pods, and Docker Containers.
  • Workspaces connect through a secure WireGuard tunnel, while developers can continue using VS Code or JetBrains.
  • Automatic shutdown of idle resources is included, which is useful for controlling the cost of long-running cloud workspaces.
  • Production deployment requires PostgreSQL 13 or later and an external access URL; the built-in database and *.try.coder.app URL are for evaluation.

Coder is therefore suitable as a unified control plane, not as a replacement for cloud, Kubernetes, or Terraform operations.

  • Self-Hosted Cloud Development Environments and AI Agents: "Workspaces are defined with Terraform"
  • Self-Hosted Cloud Development Environments and AI Agents: "EC2 VMs, Kubernetes Pods, Docker Containers, etc."
  • Self-Hosted Cloud Development Environments and AI Agents: "automatically shut down when not used"
  • Quickstart: "For a production deployment, add a PostgreSQL database (version 13 or later) and an external access URL"
coder server --postgres-url <url> --access-url <url>
Not stated in the README:The README does not specify control-plane CPU, memory, database capacity, or concurrency limits for many teams or workspaces.;The README does not state how much a Terraform template must change when reused across EC2, Kubernetes, and Docker.
Yes Our team is standardized on VS Code and JetBrains Toolbox. If workspaces run in Kubernetes Pods or Docker containers, can Coder preserve our existing IDE entry points?
For: A developer-platform engineer who provisions remote environments for VS Code and JetBrains Toolbox users without requiring a browser IDE migration

Yes. The README explicitly provides a VS Code extension and a JetBrains Toolbox plugin, and it supports development environments built on Docker, Kubernetes, and OpenShift.

  • The official VS Code Extension can open any Coder workspace with a single click.
  • The JetBrains Toolbox Plugin offers the same one-click workspace entry point, covering both existing IDE families.
  • The Dev Containers integration supports devcontainer.json on Docker, Kubernetes, and OpenShift.
  • Remote access relies on a secure WireGuard tunnel, so the actual experience still depends on network policy, proxy configuration, and latency; the README provides no performance figures.

If the requirement is to keep existing desktop IDEs rather than move to browser-based editing, Coder fits the integration model.

  • Integrations / Official: "VS Code Extension: Open any Coder workspace in VS Code with a single click"
  • Integrations / Official: "JetBrains Toolbox Plugin: Open any Coder workspace from JetBrains Toolbox with a single click"
  • Integrations / Official: "Build development environments using `devcontainer.json` on Docker, Kubernetes, and OpenShift"
  • Self-Hosted Cloud Development Environments and AI Agents: "connected through a secure Wireguard® tunnel"
coder server
Not stated in the README:The README does not specify minimum supported versions for the VS Code extension or JetBrains plugin.;The README provides no performance data for high-latency networks, large repository indexing, or debugging.
It depends We plan to modify Coder's Go and TypeScript code and provide the internal development platform to employees over a network. Could GNU AGPLv3 affect this integration model?
For: An open-source project lead who plans to embed Coder into an internal platform, modify its Go control plane or TypeScript UI, and provide the service to employees over a network

It depends. The project is licensed under GNU AGPLv3, and your network-based service and modification model may create source-code availability obligations that require legal review. The answer depends on the integration boundary.

  • The project data identifies the license as GNU Affero General Public License v3.0.
  • The README describes Coder as a self-hosted platform, which permits operation on company infrastructure but does not remove license obligations.
  • Go is the dominant implementation language and TypeScript is also a substantial part of the codebase; directly modifying the control plane or UI may involve covered code.
  • The README does not explain the compliance boundary for internal network services, closed-source extensions, derivative versions, or plugins.

Running an unmodified instance is not automatically the same as distributing software, but modifying it and offering it over a network should be reviewed by counsel before deployment, including source-code provision, modification disclosure, and module separation.

  • Project data: "license": "GNU Affero General Public License v3.0"
  • Project data: Go 32,490,987; TypeScript 9,734,073
  • README opening: "Coder is a self-hosted platform for cloud development environments and AI coding agents"
  • README / Contributing: "New contributors are always welcome"
Not stated in the README:The README does not explain the exact AGPLv3 boundary for internal employee use, modified network versions, closed-source plugins, or independent templates.;The README does not state whether commercial support or Premium features carry different licensing terms.
It depends Our development environments are in a restricted or air-gapped network. We need a private VS Code extension marketplace and self-hosted models. Can Coder meet both constraints?
For: An infrastructure engineer operating an isolated or air-gapped enterprise development platform that needs a private VS Code extension marketplace and self-hosted models

It depends. The README explicitly provides a self-hosted extension marketplace for restricted or air-gapped networks and supports self-hosted models, but it does not promise that every component required for a fully offline deployment works without additional preparation.

  • The official Self-Hosted VS Code Extension Marketplace is described as working in restricted or airgapped networks and integrating with code-server.
  • Self-hosted models are included in the supported model options, so the model service does not have to be Anthropic, OpenAI, Google, or Bedrock in the public cloud.
  • Coder itself is self-hosted and connects workspaces through WireGuard, but the deployment’s network topology and outbound dependencies must be verified.
  • Production deployment requires PostgreSQL 13+ and an external access URL; the README does not explain how these requirements are implemented in a fully isolated network.

The fit is strong for restricted networks, but a completely disconnected deployment requires validation of packages, images, model serving, and update procedures.

  • Integrations / Official: "A private extension marketplace that works in restricted or airgapped networks"
  • Self-Hosted Cloud Development Environments and AI Agents: "Bring any model (Anthropic, OpenAI, Google, Bedrock, self-hosted)"
  • Self-Hosted Cloud Development Environments and AI Agents: "connected through a secure Wireguard® tunnel"
  • Quickstart: "For a production deployment, add a PostgreSQL database (version 13 or later) and an external access URL"
Not stated in the README:The README does not list the offline images, packages, license checks, or upgrade process required for a fully air-gapped installation.;The README does not state whether self-hosted models require GPUs, which inference frameworks are supported, or what minimum resources are needed.

✨ Highlights

  • Terraform defines EC2, Kubernetes Pod, and Docker environments
  • Coder Agents runs the AI coding loop on your infrastructure
  • Supports Anthropic, OpenAI, Google, Bedrock, and self-hosted models
  • WireGuard connects workspaces and idle resources shut down automatically
  • Production deployment requires PostgreSQL 13 or later

🔧 Engineering

  • Terraform templates describe workspace infrastructure across EC2, Kubernetes, and Docker
  • Coder Agents runs in the control plane, with no LLM API keys in workspaces
  • AI Gateway centralizes authentication, auditing, and cost controls for AI tooling
  • The Quickstart uses coder server to create the first workspace at localhost:3000

⚠️ Risks

  • Production deployment needs PostgreSQL 13+ and an external access URL
  • Workspaces depend on infrastructure backends such as EC2, Kubernetes, or Docker
  • GNU AGPLv3.0 may affect compliance decisions for integration and distribution
  • The README gives no production sizing, only points to validated architectures

👥 For who?

  • Teams managing EC2, Kubernetes, or Docker development environments with Terraform
  • Teams running Claude Code, Codex, or OpenCode on their own infrastructure
  • Organizations needing centralized model authentication, audit logs, and AI cost controls
  • Developers connecting to remote workspaces through VS Code or JetBrains Toolbox