AgentRegistry v0.4.0: Declarative Control Plane
AgentRegistry v0.4.0 introduces a new declarative architecture for registering, composing, and deploying agentic infrastructure.
Important
v0.4.0 is not backward compatible with v0.3.x.
There is no supported in-place upgrade. v0.4.0 requires a fresh installation, a new database, new ar.dev/v1alpha1 manifests, and updated CLI and API integrations.
If you need compatibility with existing v0.3.x data or workflows, continue from v0.3.3.
Highlights
Declarative resources
Agents, MCP servers, Skills, Prompts, Plugins, Models, Runtimes, and Deployments now share a Kubernetes-style resource format:
apiVersion: ar.dev/v1alpha1
kind: Agent
metadata:
name: summarizer
tag: v1
spec:
# ...Resources are managed through a unified workflow:
arctl apply -f resource.yaml
arctl get agents
arctl delete -f resource.yamlExpanded catalog
AgentRegistry now catalogs six versioned artifact kinds:
- Agents
- MCP servers
- Skills
- Prompts
- Plugins
- Models
Plugins and Models are new in v0.4.0. Agents can compose references to MCP servers, plugins, skills, and prompts. Deployments select the Runtime, harness, and Model used to execute them.
Controller-driven deployments
Deployments are now reconciled asynchronously from declarative desired state.
The new controller resolves dependencies, reacts to relevant changes, avoids unnecessary runtime operations, reports status conditions, manages teardown, and tracks workloads discovered outside AgentRegistry.
Providers have been replaced by Runtime resources backed by focused local and Kubernetes adapters.
Redesigned MCP support
Remote and deployable MCP servers now use a single MCPServer resource.
Deployable packages use a structured origin, launch, and transport contract supporting npm, PyPI, and OCI artifacts.
AgentRegistry can also expose an opt-in, read-only implementation of the official MCP Registry v0.1 API. The built-in registry MCP server received additional tools, health reporting, and authorization support.
Improved developer experience
The new arctl init, build, run, apply, get, delete, pull, and wait commands provide a consistent workflow across resource types.
Built-in project frameworks include ADK Python, FastMCP Python, and MCP Go, with support for watch mode, MCP inspection, dynamic MCP resolution, and additional IDE configuration.
Operations
The Helm chart now provides explicit bundled or external PostgreSQL configuration, Secret-based database credentials, configurable bundled storage classes, and improved release-namespace handling.