Back to blog
March 12, 20265 min readEverstack Team

Introducing Everstack: The AI Infrastructure Platform

Today we are launching Everstack in alpha - a self-hosted platform that unifies AI gateway routing, agent orchestration, workflow studios, sandboxed execution, and full-stack observability into a single deployable stack.

LaunchSelf-HostedAI Infrastructure

Why We Built Everstack

Building production AI applications today means stitching together a dozen services: an LLM proxy here, an agent framework there, a separate vector database, another observability pipeline, yet another sandbox runtime. Each one has its own deployment model, its own auth surface, and its own failure modes. Teams spend more time on infrastructure plumbing than on the AI product itself.

We built Everstack to fix that. One platform, one deployment, everything you need to go from prototype to production.

What Ships in Alpha

AI Gateway

A unified proxy that routes requests across 200+ models from OpenAI, Anthropic, Google, Mistral, Cohere, and more. Built-in semantic caching, rate limiting, load balancing, and automatic failover - all configurable through YAML or the dashboard.

AI Gateway - routing, caching, and failover across providers
AI Gateway - routing, caching, and failover across providers

Agent Runtime

Deploy autonomous agents with tool use, multi-step reasoning, and persistent memory. Agents support human-in-the-loop approval workflows, sandboxed code execution, and hierarchical delegation through our job system.

Agent Runtime - tool use, reasoning, and human-in-the-loop workflows
Agent Runtime - tool use, reasoning, and human-in-the-loop workflows

Workflow Studio

A visual builder for designing multi-step AI pipelines. Drag and drop LLM nodes, conditional branches, and tool integrations. Every workflow is versioned and auditable.

Workflow Studio - visual pipeline builder with drag-and-drop nodes
Workflow Studio - visual pipeline builder with drag-and-drop nodes

Sandboxed Execution

Agents and workflows can execute arbitrary code in isolated sandbox environments backed by Docker or Kubernetes. Each sandbox gets its own filesystem, shell access, and configurable network policies.

Sandboxed Execution - isolated code environments with filesystem and shell access
Sandboxed Execution - isolated code environments with filesystem and shell access

Vector Memory

Managed vector collections powered by pgvector for RAG, semantic search, and long-term agent memory. Create collections, insert documents, and query with a single API call.

Vector Memory - collections, semantic search, and agent memory
Vector Memory - collections, semantic search, and agent memory

Full-Stack Observability

OpenTelemetry-native tracing and logging backed by ClickHouse. Every LLM call, agent turn, and workflow step is traced end-to-end with cost attribution, latency breakdowns, and token usage.

Observability - distributed tracing, cost attribution, and latency breakdowns
Observability - distributed tracing, cost attribution, and latency breakdowns

Vault

Centralized API key management with rotation policies, spend limits, and per-key access controls. Manage provider keys and application keys from a single interface.

MCP Gateway

A built-in Model Context Protocol gateway that lets agents discover and invoke MCP-compatible tool servers at runtime. Register MCP servers through the dashboard or API, and any agent can call their tools without manual wiring. This turns Everstack into a universal bridge between your agents and the growing ecosystem of MCP tool providers.

S3-Compatible Storage

Every sandbox and agent session has access to an S3-compatible object store for file uploads, artifacts, and persistent data. Use any S3 client library or the built-in SDK methods to read and write objects. Storage is scoped per workspace with configurable retention policies and size limits.

Evaluations

Ship with confidence using built-in evaluation pipelines. Define test datasets, run them against your agents or gateway configurations, and score results with LLM-as-judge, regex, code, or custom scorers. Evaluations support scheduled runs, regression detection, and side-by-side comparison across model versions - all from the dashboard or API.

Evaluations - datasets, scoring, and regression detection
Evaluations - datasets, scoring, and regression detection

Edge Functions

Lightweight serverless functions that run at the gateway layer. Use edge functions to transform requests and responses, inject context, enforce policies, or add custom routing logic - all without touching your application code. Functions execute in isolated V8 runtimes with sub-millisecond cold starts.

Self-Hosted by Default

Everstack is designed to run on your infrastructure. Deploy with Docker Compose for development or Helm charts for production Kubernetes clusters. Your data, your models, your rules.

# Get running in under 5 minutes
git clone https://github.com/everstacklabs/everstack
cd everstack
docker compose up

SDKs and CLI

The alpha ships with official SDKs for TypeScript/Node.js, Python, and Go - covering gateway chat, embeddings, agents, sandboxes, memory, datasets, and evaluations. A full CLI is included for server management and sandbox operations.

What Comes Next

This is an alpha release. We are shipping fast and iterating based on feedback. Here is what is on the immediate roadmap:

  • Apple Containerization Backend - Sub-second sandbox VMs on macOS Apple Silicon
  • Advanced Guardrails - PII detection, content filtering, and policy enforcement at the gateway layer
  • Multi-Tenant Workspaces - Isolated environments with per-tenant quotas, billing, and access controls
  • Prompt Registry - Versioned prompt management with A/B testing and rollback

Get Involved

Everstack is built for teams that want a unified, self-hosted AI infrastructure platform. We welcome bug reports and feature requests.

We are excited to see what you build with Everstack. Deploy it, break it, and tell us what you think.