Eve: Your always-on SRE assistant for Slack

Eve is a next-generation Kubernetes operations agent built in Go. It functions as an intelligent MCP (Model Context Protocol) Proxy that connects your Slack workspace to a local LLM and a modular ecosystem of MCP servers.

Instead of hardcoding tools, Eve acts as a Supervisor Agent that dynamically discovers capabilities from external providers (Kubernetes, GitHub, Argo, etc.) and orchestrates them through natural language.

---

🏗 Architecture

Eve sits at the intersection of your communication (Slack), your "brain" (Local LLM), and your tools (MCP Servers).

1. Core Orchestration

Eve acts as a secure gateway between your Slack workspace and technical tools.

graph TD
    User([User in Slack]) <-->|Socket Mode| Eve[Eve Bot/Proxy]
    Eve <-->|Tool Calls| LLM[LLM: Qwen3-Coder]

    subgraph "MCP Ecosystem"
        Eve <-->|JSON-RPC| K8s[K8s MCP Server]
        Eve <-->|JSON-RPC| GH[GitHub MCP Server]
        Eve <-->|JSON-RPC| Argo[Argo MCP Server]
        Eve <-->|JSON-RPC| AWS[AWS MCP Servers]
    end

    K8s <-->|API Server| Cluster[...]
    GH <-->|GitHub API| GitHub[...]
    AWS <-->|Cloud API| Cloud[...]

Loading

2. Memory & Context System

Eve maintains both short-term conversational context and long-term technical experience.

graph TD
    Agent[Eve Agent] --> Store[Memory Store]
    Store --> SQLite[(SQLite: Short-term Memory)]
    Store --> Qdrant[(Qdrant: Long-term Memory)]
    Store --> Embedder[Embedder: Ollama/OpenAI]

    SQLite ---|Conversation History| Agent
    Qdrant ---|Semantic Retrieval| Agent

Loading

---

📸 Demo

Eve operates in two distinct modes, automatically detecting which mode to use based on your message:

🔧 SRE Mode (Infrastructure & Operations)

When you ask about pods, deployments, logs, or any infrastructure topic, Eve immediately calls the appropriate MCP tools to fetch real data. No hallucination, no fake outputs.

"fft-os 네임스페이스 pod 상태 확인해줘" → Calls pods_list_in_namespace tool

💬 Chat Mode (Casual Conversation)

For greetings, general questions, or casual chat, Eve responds naturally without calling any tools. It matches your language (Korean/English) automatically.

"안녕! 오늘 기분 어때?, grpc가 뭐야?" → Friendly conversation, no tools

---

🚀 Key Features

• Philosophy: "Don't Reinvent the Wheel": Eve doesn't contain domain-specific logic. It focuses on the orchestration gateway, while domain logic is outsourced to standardized MCP servers.
• Agentic Supervisor: Uses a ReAct-style loop to handle complex multi-step operations (e.g., "Find the failing pod, check its logs, and create a GitHub issue if it's an OOMKill").
• Agentic Delegation: Minimizes direct command execution using slash commands in favor of delegating complex tasks to the agent via MCP and dynamic toolsets, allowing Eve to orchestrate actions autonomously.
• Dynamic Tool Discovery: At startup, Eve performs a handshake with all configured MCP servers to list and register their tools.
• Memory: Maintains both short-term conversational context and long-term technical experience.
• Zero-Ingress Security: Uses Slack Socket Mode for outbound-only connections.
• Local-First AI: Optimized for local LLM execution. Default-configured for Qwen3-Coder via llama-cpp or Ollama.
• Sidecar Optimized: Designed to run in Kubernetes with MCP servers as sidecars for low latency and shared RBAC.

---

🛠 Quick Start

1. Configuration

Eve can be configured via environment variables or a standard mcp.json file.

.env Configuration:

# Slack
SLACK_APP_TOKEN=xapp-...
SLACK_BOT_TOKEN=xoxb-...

# LLM (OpenAI-Compatible endpoint)
LLM_PROVIDER=openai
LLM_BASE_URL=http://qwen.home.lab:8003
LLM_MODEL=qwen3-coder

# MCP Servers (Comma-separated URLs)
MCP_SERVERS=http://localhost:8080,http://localhost:8081

mcp.json Configuration:

{
  "mcpServers": {
    "kubernetes": { "url": "http://localhost:8080" },
    "github": { "url": "http://localhost:8081" },
    "aws-billing": { "url": "http://aws-billing:8083" }
  }
}

2. Deployment (Sidecar Pattern)

The recommended way to deploy Eve is with the official kubernetes-mcp-server as a sidecar:

# manifests/base/deployment.yaml
containers:
  - name: eve
    image: harbor.home.lab/restack/eve:latest
    env:
      - name: MCP_SERVERS
        value: "http://localhost:8080"

  - name: mcp-kubernetes
    image: quay.io/podman/kubernetes-mcp-server:latest
    args: ["--port=8080"]

---

🤖 Recommended LLM

Eve is designed to work with models that have strong Tool Calling (Functional Calling) capabilities.

• Recommended: qwen3-coder (Excellent at precise tool selection and SRE tasks).
• Alternative: qwen2.5:14b or llama3.1:8b via Ollama.

---

📦 CI/CD

Eve includes a production-ready GitHub Actions workflow (.github/workflows/build-image.yml) that:

1. Builds a minimal Go binary into a scratch-based Docker image.
2. Pushes to your private registry (Harbor).
3. Automatically updates your GitOps manifests in your homelab repository.

---

🔧 Extensibility

Adding a new capability to Eve is as simple as adding a new URL to your MCP_SERVERS list. Whether it's a Jira agent, a database explorer, or a custom internal tool, as long as it speaks MCP, Eve can use it.

# Example: Adding a Jira MCP server
export MCP_SERVERS="http://k8s-mcp:8080,http://jira-mcp:8080"

---

⚖️ License

MIT © 2026 Restack / Eve Team