Everstack is an open-source AI infrastructure platform for building and operating production AI applications. It combines an OpenAI-compatible AI gateway, agent runtime, isolated sandboxes, memory, MCP, workflows, evaluations, guardrails, and observability in one control plane.

Is Everstack an AI gateway or an agent platform?

It is both, plus the infrastructure around them. The gateway handles multi-provider routing, fallbacks, caching, rate limits, and keys. The runtime adds stateful agents, tools, memory, workflows, isolated execution, evaluations, and end-to-end traces, so teams do not have to join separate platforms themselves.

Can I self-host Everstack?

Yes. Deploy Everstack on your own infrastructure with Docker or Kubernetes. You operate its storage and runtime, but external model providers and tools still receive the data your application sends to them. Features and limits depend on the edition and configuration; self-hosting alone does not make a workload air-gapped.

Which AI providers are supported?

Everstack supports OpenAI, Anthropic, Google Gemini, Azure OpenAI, AWS Bedrock, Cohere, Mistral, DeepSeek, Hugging Face, Qwen, MiniMax, Moonshot, Ollama, OpenRouter, and other OpenAI-compatible endpoints.

How does Everstack isolate agent code execution?

Agent tasks run inside managed sandbox boundaries with explicit filesystem, network, resource, and lifecycle controls. Everstack keeps the runtime behind that stable sandbox contract so applications do not depend on the underlying isolation technology.

How do evaluations and observability work together?

Everstack records model calls, tool use, sandbox execution, latency, tokens, cost, and outcomes as connected traces. Those production traces can feed datasets, LLM-as-judge scorers, code scorers, human review queues, and regression checks, closing the loop between what shipped and what should improve next.

Does Everstack lock me into a framework or model provider?

No. Everstack exposes OpenAI-compatible APIs, native TypeScript, Python, and Go SDKs, Connect RPC services, and OpenTelemetry data. It works with common agent frameworks and model providers, so applications can change models or orchestration code without replacing the infrastructure layer.

Everstack is an open-source AI infrastructure platform for building and operating production AI applications. It combines an OpenAI-compatible AI gateway, agent runtime, isolated sandboxes, memory, MCP, workflows, evaluations, guardrails, and observability in one control plane.

Is Everstack an AI gateway or an agent platform?

It is both, plus the infrastructure around them. The gateway handles multi-provider routing, fallbacks, caching, rate limits, and keys. The runtime adds stateful agents, tools, memory, workflows, isolated execution, evaluations, and end-to-end traces, so teams do not have to join separate platforms themselves.

Can I self-host Everstack?

Yes. Deploy Everstack on your own infrastructure with Docker or Kubernetes. You operate its storage and runtime, but external model providers and tools still receive the data your application sends to them. Features and limits depend on the edition and configuration; self-hosting alone does not make a workload air-gapped.

Which AI providers are supported?

Everstack supports OpenAI, Anthropic, Google Gemini, Azure OpenAI, AWS Bedrock, Cohere, Mistral, DeepSeek, Hugging Face, Qwen, MiniMax, Moonshot, Ollama, OpenRouter, and other OpenAI-compatible endpoints.

How does Everstack isolate agent code execution?

Agent tasks run inside managed sandbox boundaries with explicit filesystem, network, resource, and lifecycle controls. Everstack keeps the runtime behind that stable sandbox contract so applications do not depend on the underlying isolation technology.

How do evaluations and observability work together?

Everstack records model calls, tool use, sandbox execution, latency, tokens, cost, and outcomes as connected traces. Those production traces can feed datasets, LLM-as-judge scorers, code scorers, human review queues, and regression checks, closing the loop between what shipped and what should improve next.

Does Everstack lock me into a framework or model provider?

No. Everstack exposes OpenAI-compatible APIs, native TypeScript, Python, and Go SDKs, Connect RPC services, and OpenTelemetry data. It works with common agent frameworks and model providers, so applications can change models or orchestration code without replacing the infrastructure layer.

New in the catalogJev 1.13 + Kev (configured checkpoint)

Open-source AI infrastructureCloud or self-hosted

Run AI applications with control over models, execution, and cost.

Route model requests, run tool-using agents, and investigate failures in one platform. Start with the OpenAI-compatible gateway, then add the runtime services your application needs. Self-host Everstack or use Everstack Cloud.

Start with $5 creditSelf-host Everstack

One-time starter credit per account on Everstack Cloud. Expires 30 days after signup.

01OpenAI-compatible API
02Agents, workflows, and sandboxes
03Guardrails and evaluations
04OpenTelemetry-native

New in the model catalog

v2.10.0 · 2026-09-29

Read the catalog changelog

Jev 1.13
Input
Output
Kev (configured checkpoint)
Input
Output

Catalog additions. Availability depends on your provider access, gateway version, and the model's supported API.

Browse all models

The product

See how the pieces work together.

Explore the gateway and agent runtime through these interactive examples. The Sites tab previews planned publishing capabilities. All panels use sample data.

Observability

Monitoring

Analytics

Segments

Logs

Live

Search Logs...

14dLast 14 days

2026-05-15 17:33:18CHATOpenAIgpt-5.2-codex3.57s5.1k↑ / 97↓$0.00483.6s TTFT-corr_01k8n4z2p9xq..."test"

2026-05-15 16:48:02CHATAnthropicclaude-opus-4.72.84s1.2k↑ / 231↓$0.00711.9s TTFT76.4 t/scorr_01k8n3w7m2rd..."can you review this deployment config..."

2026-05-15 15:12:44CHATGooglegemini-2.5-pro4.21s2.8k↑ / 418↓$0.00392.7s TTFT64.8 t/scorr_01k8mzx9q4la..."summarize the trace errors from last run..."

2026-05-15 13:57:19AGENTOpenAIgpt-5.2-codex30.0s3.4k↑ / 0↓$0.0011--corr_01k8mt2c7vne...upstream_timeout: provider did not respond within 30s

2026-05-15 12:22:06CHATMistralcommand-a1.92s743↑ / 118↓$0.00121.4s TTFT88.7 t/scorr_01k8mpq6t1bz..."rewrite this onboarding copy..."

2026-05-14 18:41:55TOOLAnthropicclaude-sonnet-4.55.76s1.9k↑ / 384↓$0.00562.4s TTFT71.3 t/scorr_01k8kz93fh2x..."call github.searchRepos with filters..."

2026-05-14 16:08:31CHATOpenAIgpt-5.1-mini1.34s612↑ / 89↓$0.00060.8s TTFT103.5 t/scorr_01k8kq1zr5pa..."how about now"

2026-05-14 14:33:07AGENTMistralmistral-large-latest4.89s2.1k↑ / 297↓$0.00282.9s TTFT59.6 t/scorr_01k8kjc4y8wd..."generate the migration plan for shared tenancy..."

2026-05-14 14:31:52AGENTOpenRouteropenrouter/auto3.42s1.8k↑ / 243↓$0.00311.7s TTFT71.1 t/scorr_01k8kh9r2v6t..."generate the migration plan for shared tenancy..."

2026-05-14 12:58:49CHATOpenRouteranthropic/claude-3.7-sonnet0.41s1.4k↑ / 0↓$0.0000--corr_01k8kdm8r6pb...rate_limit_exceeded: provider returned HTTP 429

2026-05-14 11:37:26TOOLOpenRoutermeta-llama/llama-3.3-70b-instruct2.66s986↑ / 174↓$0.00181.5s TTFT82.5 t/scorr_01k8kadc3tme...tool_execution_failed: github.searchRepos returned 403

2026-05-14 11:49:23CHATAnthropicclaude-opus-4.73.18s932↑ / 156↓$0.00422.1s TTFT80.2 t/scorr_01k8k9m7v6sc..."ok can you clone the es-runtime repo..."

2026-05-14 10:20:52AGENTOpenAIgpt-5.2-codex3.88s787↑ / 144↓$0.00213.0s TTFT46.3 t/scorr_01k8k72bd9he..."do you have access to the github repo..."

production control plane

live view

healthy

Gateway

routing · fallback · cache

4 running

Agents

sessions · tools · memory

isolated

Sandboxes

docker · firecracker · k8s

94% pass

Evaluations

judges · scorers · reviews

connected

MCP + Workflows

tools · DAGs · approvals

live

Observability

traces · cost · alerts

identity, policy, state, and telemetry stay attached

Route, evaluate, and trace every model request from one production control plane.

Explore platform docs

01The production story

One workload. Four decisions. One continuous system.

A production AI workload has to choose a model, execute real work, cross policy and quality checks, then return evidence for the next release. Everstack keeps those handoffs connected.

01Route the request
02Execute the work
03Verify the boundary
04Learn from production

An operator routing model traffic through an illustrated infrastructure switchboard

01Route the request

Start with one request, not one provider.

Point existing OpenAI clients at one endpoint. Everstack chooses the provider, applies tenant controls, handles fallback, and records the economics without pushing that logic into your application.

One OpenAI-compatible contract
Health-aware routing and fallback
Tenant-scoped keys and limits
Traceable latency, tokens, and cost

Explore the gateway Compare models

Hosted, cloud, and local models stay behind the same request contract.

Explore the model catalog

An illustrated operator arranging a clear sequence of connected AI workflow stations

02Execute the work

Give your application tools and a place to run.

Use agents when the path must adapt, workflows when the path must stay explicit, and isolated sandboxes when the workload needs code, files, browsers, or long-running processes.

Agents for adaptive decisions
Workflows for explicit control
Isolated Docker, Firecracker, or Kubernetes
Persistent sessions, memory, and artifacts

Agent runtime Workflows Sandboxes

An illustrated checkpoint inspecting AI requests before they enter a protected system

03Verify the boundary

Make policy and quality part of execution.

Check inputs before model spend, pause sensitive actions for approval, inspect outputs before delivery, and compare every change against evidence instead of intuition.

Input and output guardrails
Human approval boundaries
LLM judges and code scorers
Datasets, baselines, and release gates

Guardrails Evaluations

An illustrated operator observing AI workload signals from a night-time monitoring station

04Learn from production

Let every run explain the next change.

Follow the model choice, tool calls, sandbox operations, approvals, evaluation scores, latency, tokens, and cost as one connected trace, then use that evidence to improve the next release.

End-to-end OpenTelemetry traces
Token and cost attribution
Trace-linked evaluation scores
Production evidence for the next release

Observability Evaluations

Your first request. Your next application.

Choose how you operate Everstack

Use Everstack Cloud for managed infrastructure, or deploy the Community Edition on your own infrastructure. Features and limits depend on your edition and plan. External model providers and tools still receive the data you send to them.

Everstack Cloud

Your AI stack, managed.

Run the services your application needs without managing the infrastructure.

Includes $5 in starter usage credit, valid for 30 days.

Start building

No infrastructure to manage. Add services as your application grows.

Start with one model request

Connect your provider, send a request through the OpenAI-compatible gateway, and inspect its trace. Add fallback routes and usage limits as your application grows.

Your applicationEverstack gatewayYour provider

Follow the gateway quickstart

Add execution when your application needs it

Run tool-using agents with persistent sessions, connect MCP tools, and execute code in isolated sandboxes. Investigate model and tool behavior through OpenTelemetry traces.

Explore the agent runtime