Roadmap¶

sikifanso is an AI agent infrastructure platform. Development follows a phased approach, building from cluster foundations toward AI-powered operations.

Phase 0: Foundation & Identity -- shipped¶

Replaced the original homelab catalog with AI agent infrastructure tools. 17 curated tools across 7 categories: gateway, observability, guardrails, RAG, runtime, models, and storage. Retained the existing k3d bootstrapping, GitOps catalog system, doctor checks, snapshots, TUI browser, dashboard, and dual-track app model.

Phase 1: Agent Cluster Profiles -- shipped¶

sikifanso cluster create --profile <name> enables a pre-defined set of catalog apps. Profiles are composable: --profile agent-dev,rag enables the union of both. Available profiles: agent-minimal, agent-full, agent-dev, agent-safe, rag.

See Profiles for details.

Phase 2: Agent Isolation & Network Policies -- shipped¶

sikifanso agent create <name> scaffolds a namespace with resource quotas, network policies, and service account. Cilium NetworkPolicies enforce default-deny egress, allowlisted data layer access, no cross-agent traffic, and no Kubernetes API access.

See Agent Sandboxes for details.

Phase 3: MCP Server Interface -- shipped¶

sikifanso mcp serve exposes cluster operations as MCP tools (stdio transport). 25 tools across cluster management, catalog, agents, ArgoCD, Kubernetes, and health checks. Any MCP-compatible client can manage the cluster.

See MCP Server for details.

Phase 4: AG-UI Protocol Integration¶

Stream agent reasoning and actions to terminal/web UI in real-time.

• AG-UI SSE endpoint in the dashboard server
• Event flow: RUN_STARTED -> STEP_STARTED/FINISHED -> TOOL_CALL_* -> TEXT_MESSAGE_* -> STATE_SNAPSHOT/DELTA
• Terminal renderer for structured output
• Web dashboard live activity feed
• sikifanso agent watch <name> for live agent activity streaming

Phase 5: AI-Powered Operations¶

Natural language cluster operations powered by MCP tools and AG-UI streaming.

• sikifanso ai "<prompt>" -- natural language cluster operations via Claude API with tool use
• AI uses the same MCP tools that external agents use (Phase 3)
• Reasoning streamed via AG-UI (Phase 4)
• Intelligent troubleshooting: sikifanso ai "why is langfuse unhealthy"
• AI-driven catalog: sikifanso ai "I need to run agents with guardrails and cost tracking"

Phase 6: Advanced Features¶

Production hardening, community, ecosystem.

• Shareable agent cluster profiles (OCI artifacts / git refs)
• Community catalog contributions (curated AI tool definitions)
• Multi-cluster agent federation
• GPU scheduling support (Ollama/vLLM with NVIDIA GPUs)
• Cloud cluster support (bootstrap EKS/GKE with the same catalog)
• Agent marketplace: pre-built agent templates that run on sikifanso clusters

Phase dependencies¶

Phase 0 (Foundation) -- DONE
  +---> Phase 1 (Profiles) -- DONE
  +---> Phase 2 (Isolation) -- DONE
           +---> Phase 3 (MCP Server) -- DONE
                    +---> Phase 4 (AG-UI)
                             +---> Phase 5 (AI Ops)
                                      +---> Phase 6 (Advanced)

Other ideas¶

These are directions worth exploring, not committed to any phase:

• Remote GitOps repos -- support remote git repos as the gitops source for collaboration and CI/CD. See Remote GitOps.
• Additional CNI options -- Calico, Flannel, or bring-your-own via a --cni flag.
• Cluster templates -- shareable templates defining node count, resource limits, pre-installed apps, and network policies.
• Multi-node topologies -- HA control plane, agent nodes for workload isolation.
• Terraform / OpenTofu integration -- a provider or module that wraps sikifanso.