What is Evenfire?
Evenfire is an open-source platform for running LLM agents inside your own Kubernetes cluster. Agents talk to people over a desktop app, Telegram, Microsoft Teams, Slack and email, call tools over MCP, run shell commands, make HTTP requests and drive a browser. Commands, outbound HTTP and browser control pause for a person to approve from the channel before they run.
Agents, connectors, channels, workflows, policies and plugins are Kubernetes custom resources, nine in all, so the fleet lives in version control. Each agent runs in its own pod in a namespace that starts with no network access; connectors are opened per agent, and token budgets can block or warn by scope. It ships with a Control UI for admins, an Electron desktop app and a profile page for invitations, and works with 24 model providers on your own keys.
The code is TypeScript under MPL-2.0 and is in beta at v0.12.0. The single-tenant platform is complete in the repository; Evenfire, the company, sells a hosted multi-tenant service with support and an SLA, at prices it does not publish. The code uses the internal name clerum.
Is Evenfire worth it?
Evenfire takes agent safety as an infrastructure problem rather than a prompt problem, and that is its strongest argument. Each agent gets its own pod, runtime namespaces start with no connectivity, connectors are opened per agent with short-lived scoped tokens, and risky tool calls suspend until someone approves them in Slack, Teams or Telegram. Frameworks such as LangGraph leave those controls to the application; here the platform enforces them.
The defaults deserve a careful read. Agent pods may reach the public internet on ports 80 and 443 so they can call model APIs and the approval-gated HTTP tool, while private and metadata ranges stay blocked. The connector image allowlist is off unless you set CONTROL_API_ENFORCE_IMAGE_ALLOWLIST=true, and the minikube overlay switches on the ChatGPT and Grok subscription sign-ins that the base deployment leaves off. Production clusters need a NetworkPolicy-capable network plugin such as Calico, or the isolation does not exist.
The operating cost is real. The local trial needs Docker with at least 10 GB of memory and six CPUs, minikube, kubectl, Node.js 24, Python and Ruby, and takes about 15 minutes on a cold cache. Production means adapting the kustomize overlays to your own cluster, managing signing keys and secrets, and running more than twenty services. A team without Kubernetes skills will spend its first weeks on the platform, not the agents.
Maturity is the open question. The project is four months old, with 575 stars, releases up to v0.12.0 on 28 October and commits as recent as 5 November, but also 793 open issues and pull requests. The recent ones are specific and sometimes serious: GFS file tokens outliving a logout, an account recreated by replaying an old invitation, context compaction failing on OpenAI reasoning models. Filing them in public is a good sign; the maintainers say APIs and CRDs may change before 1.0.
Compared with OpenClaw, which puts a personal assistant on your phone in an afternoon, Evenfire is aimed at platform teams that need many agents, many users and an audit trail. Dify gets a team to a working agent faster but enforces far less around what that agent may touch.
Where it earns its place
A platform team at a 300-person SaaS company with an existing EKS cluster
Let an ops agent run diagnostic commands without giving it free rein
Declare the agent and its connectors as CRDs, keep shell_exec behind the approval gate so each command waits for an on-call engineer in Slack, and set a token budget for the team. Cost is cluster capacity plus model spend on your own keys.
An IT lead at a regulated firm that cannot send data to a hosted assistant vendor
Give staff a shared assistant over Teams that stays on our infrastructure
Self-host on an internal cluster with a key for Bedrock or Vertex AI; prompts go only to that provider, shared-drive reads and writes are logged, and invitations go through the Profile UI. Support is community-only without the hosted plan.
A developer evaluating agent platforms on a laptop
See whether the approval and isolation model fits before proposing it
Run make minikube-setup with one model key; a seeded chatllm agent answers in the desktop app after roughly 15 minutes of image pulls. Needs 10 GB of memory free for Docker.
Before you commit
- Setup
- Local trial: clone, set ADMIN_PASSWORD and one model key in .env, run make minikube-setup (about 15 minutes cold) with Docker at 10 GB of memory and six CPUs, minikube 1.30+, kubectl, Node 24, Python and Ruby. Production: adapt deploy/overlays to a cluster with Calico or equivalent.
- Integrations
- Telegram, Slack, Microsoft Teams, email and a desktop app as channels; MCP connectors including MongoDB and Airtable; 24 providers including OpenAI, Anthropic, Bedrock, Vertex AI, Azure OpenAI and OpenRouter; plugin registry; GitOps through CRDs.
- Data and privacy
- Self-hosted; prompts and files go only to the model provider you configure. Shared-file access is checked and audited per request. The optional shared registration hub sends invitation emails on your behalf. MPL-2.0.
- Cost at scale
- Self-hosting costs cluster capacity for 20-plus services and one pod per agent, plus model spend on your keys; token budgets cap it per scope. The hosted service's price is not published.
- Limits
- One organisation per self-hosted deployment; community support with no SLA; beta APIs; 793 open issues and pull requests including logout and invitation-replay bugs.
What’s good
- Approval gate on shell, outbound HTTP and browser actions
- Per-agent pods with default-deny networking
- Whole fleet declared as nine Kubernetes CRDs
- 24 model providers on your own keys
- Single-tenant platform complete under MPL-2.0, not feature-gated
What’s not
- Needs Kubernetes skills and a NetworkPolicy-capable cluster
- Beta: APIs and CRDs may change before 1.0
- 793 open issues and pull requests, some security-related
- Image allowlist off by default; hosted pricing not published
Best for
Platform teams already running Kubernetes who need several agents with real system access, shared across users, under approvals, budgets and audit logs.
Evenfire alternatives
Evenfire: common questions
- Is Evenfire worth it?
- Evenfire puts serious guardrails around agents that touch production systems, at the cost of running a Kubernetes platform that is still in beta. We rate it 3.6 out of 5.
- How much does Evenfire cost?
- Free under MPL-2.0 for self-hosting, one organisation per deployment. Hosted multi-tenant service with support and SLA; price not published.
- What are the downsides of Evenfire?
- Needs Kubernetes skills and a NetworkPolicy-capable cluster; beta: APIs and CRDs may change before 1.0; 793 open issues and pull requests, some security-related; image allowlist off by default; hosted pricing not published.
- Who is Evenfire best for?
- Platform teams already running Kubernetes who need several agents with real system access, shared across users, under approvals, budgets and audit logs.
- How much does Evenfire cost for a team or in production?
- Self-hosting costs cluster capacity for 20-plus services and one pod per agent, plus model spend on your keys; token budgets cap it per scope. The hosted service's price is not published.
- Is Evenfire safe for company data?
- Self-hosted; prompts and files go only to the model provider you configure. Shared-file access is checked and audited per request. The optional shared registration hub sends invitation emails on your behalf. MPL-2.0.
- What are the best Evenfire alternatives?
- The closest alternatives we have reviewed are OpenClaw, Dify and n8n.
Answers checked Nov 5, 2026.