evenfire-ai/evenfire

Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure

What it solves

Evenfire provides a secure, self-hosted platform for deploying LLM agents that can perform real actions on company infrastructure. It addresses the risk of "hallucinating" agents by implementing strict human-in-the-loop approval gates for risky actions, while ensuring data privacy by allowing users to bring their own model keys and host the entire system on their own Kubernetes cluster.

How it works

The platform is built as a suite of Kubernetes custom resources (CRDs), treating the entire agent fleet—including agents, connectors, workflows, and policies—as version-controlled configuration. It uses the Model Context Protocol (MCP) to allow agents to call tools and interact with plugins. To maintain security, it employs a "deny-all" networking approach where connectivity is only opened per agent/connector pair, and risky commands (like shell execution or outbound HTTP) are paused until a human provides a signature-verified approval.

Who it’s for

Companies and teams that need AI agents to automate workflows and interact with their internal tools, but require high levels of governance, security, and auditability that hosted assistants cannot provide.

Highlights

  • Human-in-the-loop governance: Risky actions are automatically paused for human review and approval.
  • Multi-channel support: Agents can communicate across a native desktop app, Telegram, Microsoft Teams, and Slack.
  • Infrastructure as Code: The entire platform is managed via Kubernetes CRDs, making it GitOps-friendly.
  • Model Agnostic: Supports a wide range of providers (OpenAI, Anthropic, Google, etc.) and local LLMs, preventing vendor lock-in.
  • Secure Isolation: Uses per-agent pods and pinned egress proxies to restrict network access.

Related

  • Project
    ayi-ai/nie-grassroots-logic

    An AI-ready methodology toolkit that transforms the frameworks from Nie Huihua's book on grassroots China into executable skills for AI agents to analyze power structures and local governance.

  • Project
    BlockRunAI/Franklin

    An autonomous economic AI agent that holds a USDC wallet to pay for models and tools, enabling it to perform trading, research, and coding tasks with a hard budget cap.

  • Project
    giselles-ai/giselle

    An open-source AI agent studio for building and runing agentic workflows with a visual builder and GitHub integration.

  • Project
    skygazer42/MimirQ

    An observable and replaceable enterprise RAG infrastructure that provides a controllable pipeline for document parsing, chunking, and retrieval with built-in regression testing.