Run Self-Hosted Agents in 2.33

Run AI development workflows on self-hosted infrastructure with Coder Agents in Beta

Most agent tools are tightly coupled to a single provider, leaving platform teams with fragmented workflows and no standard way to govern how agents operate across the org.

Coder Agents is a native coding agent that runs on your control plane. It enables teams to have centralized control over models, prompts, MCPs, and network-isolated workspaces, while keeping model and provider choice fully flexible. Coder Agents is available in beta, with introductory access that includes full Coder Agents Premium features and no usage-based limits through September.

• Self-hosted by default: The agent loop runs on your own infrastructure, with full support for air-gapped deployments and self-hosted models.
• Centralized governance: Manage models, system prompts, MCP servers, and agent skills from a single control plane so every team runs agents consistently.
• Model flexibility without lock-in: Bring any frontier or self-hosted LLM (Anthropic, OpenAI, Google, Bedrock, or your own endpoint) and switch providers without changing your workflows.

Learn more on our blog.

Expanded endpoint control and authentication flexibility in the AI Governance Add-on

2.33 gives platform teams more granular control over how models are accessed, tracked, and governed.

• Same provider, multiple endpoints: Admins can now configure multiple named endpoints per provider. For example, a direct connection to Anthropic alongside one routed through AWS Bedrock, or one endpoint with Zero Data Retention enabled and another without. Each endpoint has its own base URL and API key, meeting compliance and infrastructure requirements while maintaining centralized governance.
• AWS Bedrock IAM-based authentication: AWS Bedrock now supports IAM role-based authentication. Static credentials are optional, enabling organizations that rely on IAM Roles instead of IAM Users to authenticate and access Bedrock resources.

New resource: agent skill for template creation and management

Build and manage Coder templates faster with a new agent skill that encodes Coder's internal expertise directly into your coding agent workflows.

• Fits your current setup: Use it locally within your existing coding agent workflows. Simply add the skill when creating or updating templates. Browse available skills in the Coder Registry.
• Consistent quality templates at scale: Skills define reusable agent capabilities that teams can standardize on while in the Coder ecosystem.

Leverage agent modules across any template

Agent modules can now be installed and composed freely across any template, giving teams more flexibility to build and manage their environments. Available in all plans.

• More flexible templates: Admins can now include multiple agent modules (e.g., Claude Code + Codex) in a single template. They can all activate at runtime, or admins can let users choose which one they want.
• Cleaner, safer defaults: Modules install in CLI-only mode by default with AI Governance integrations as explicit opt-ins, making it easier to compose agents for interactive or agentic workflows without unintended side effects.

Breaking changes

• #24650: Terminal command confirmation: Users are now prompted to confirm before executing terminal commands from a URL. This is a security-focused change.
• #22168: Removed unnecessary api.ts calls, which may affect custom API integrations.

View the full changelog on GitHub. If you have questions or feedback, join the conversation on Discord or email us directly!