Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
Why evenfire
Build multi-channel LLM agents you own — agents that take real actions, with approvals, permissions, and audit-oriented controls built in. Self-host on your infrastructure; bring your own model keys.
Core idea (product messaging):
A model can be steered, misled, or simply wrong. So it is never trusted by default.
The problem
Once an agent can call tools (send mail, write data, open a shell, hit APIs), prompt quality alone is not enough. You need:
- Who may talk to the agent (channels, allowlists)
- What tools it may reach (connectors, network policy)
- When a human must approve (default-on gates)
- How the fleet is versioned and reviewed (declarative config)
Personal chat assistants optimize for “message me tonight.” In-process frameworks optimize for “compose agents in application code.” evenfire optimizes for governance around action when you still want a full agent runtime and multi-channel surfaces.
What evenfire optimizes for
- Human-in-the-loop by default — with the approval system on (the default),
MCP tool calls always pause for explicit approve/deny, as do the risky native
tools (
shell_exec,http_request,cron_manage,browser_open,browser_navigate); pending approvals survive restarts. - Least privilege at the network layer — a
Contextallowlists MCP servers; unlisted servers are unreachable via default-deny NetworkPolicies. - Declarative fleet — Hosts, channels, connectors, and workflows are
Kubernetes CRDs (
clerum.io). Review them like any other infra change. - Model-neutral BYOK — one interface across a broad, growing roster of providers (see Supported LLM providers); switching is a config change.
- Self-hosted data path — prompts and tool traffic stay on your infrastructure.
Who it is for
- AI / app developers building multi-channel agents they operate themselves
- Teams that need Telegram / Slack / email / desktop under shared policy
- Operators who want the full security model from day one — one command
(
make minikube-setup) runs the whole platform on a local cluster
Who it is not for (yet)
- Someone who only wants a single-process personal assistant with one-line install and no interest in self-hosted policy
- Developers who only need an in-process agent library inside an existing app (no platform control plane)
See When to use evenfire for category-level fit.
Vocabulary bridge
| You might say | In this repo |
|---|---|
| agent | Host CRD + mcp-host runtime |
| governed connector | McpServer scoped by Context |
| plugin / workflow | WorkflowRecipe (+ registry client) |
| model-neutral, BYOK | many providers, one interface; keys in your Secret |
| shared files | SharedFileSystem + GlobalFileSystem |
Next
- Quickstart — try the agent runtime locally
- Security model — four enforcement layers
- Architecture overview — services and lifecycle
- Open core: self-host vs hosted — what's in this repo vs the managed service