Resend

Everruns

Note: Everruns is under active development. Expect rapid changes and experimental features.

Everruns is an open-source platform for building, deploying, and operating durable AI agents.

Define agents and their tools, compose them into reusable harnesses, then ship them to real users through Slack, web chat, scheduled jobs, webhooks, A2A, MCP, or a plain HTTP API — backed by a Rust + PostgreSQL durable execution engine that survives restarts and scales horizontally.

Why Everruns

• Durable by default — every session is a PostgreSQL-backed workflow that survives restarts, worker crashes, and network partitions. No lost runs, no in-memory state to babysit.
• One agent, every channel — define a harness once, then publish it to Slack, web chat, A2A, webhooks, cron schedules, voice, or a plain HTTP/MCP API.
• Open and provider-neutral — implements the Open Responses spec across OpenAI, Anthropic, Gemini, and Azure; register remote MCP servers as tools. MIT-licensed and self-hostable.
• Built for production — multi-tenant orgs, fine-grained permissions, envelope-encrypted secrets, budgets, observability, and a control-plane / worker split that scales horizontally.

Architecture

graph TB
    subgraph Public["Public surface"]
        Clients["Clients<br/>SDK · CLI · App"]
        Proxy["Reverse proxy<br/>TLS · /api · /mcp"]
    end
    subgraph CP["Control plane"]
        Server["Server<br/>REST · gRPC"]
    end
    subgraph Workers["Worker pool · stateless"]
        W1["Worker"]
        W2["Worker"]
        Wn["Worker N"]
    end
    subgraph Data["State & messaging"]
        PG["PostgreSQL<br/>required · task queue + event log"]
        NATS["NATS JetStream<br/>optional · event push"]
        Valkey["Valkey<br/>optional · rate limiting"]
    end

    Clients --> Proxy --> Server
    Server <--> W1
    Server <--> W2
    Server <--> Wn
    Server --> PG
    Server -.-> NATS
    Server -.-> Valkey

Workers are stateless and hold no database credentials — they reach the control plane over gRPC only. See Architecture for the full picture.

What's inside

Build agents

• Harnesses, Agents, Capabilities, Skills — modular configuration that composes into a runtime agent. Built-in harness types for general chat, data analysis, and coding (sandbox or Daytona-backed).
• Agent versioning, blueprints, and identities — iterate safely on production agents and run unattended workloads under virtual principals.
• Capabilities library — web fetch, bash sandbox, session filesystem, session SQL DB, memory, knowledge bases, MCP, voice, and more.
• Skills registry — agentskills.io-format skills with discovery, search, and usage tracking.
• Generative UI — agents can return rich cards via OpenUI, A2UI, and inline ui://everruns/... MCP App resources.

Connect any model and any tool

• LLM providers: OpenAI (Responses + Chat Completions), Azure OpenAI, Anthropic, Gemini, plus llmsim for tests. Model resolution is layered: message → session → agent → system default.
• MCP, both ways — register remote MCP servers as virtual capabilities (auto-discovered and namespaced), and expose your own deployment's MCP endpoint with OAuth 2.1 so other agents can call yours.
• Integrations: Docker, Daytona, E2B, Deno, Browserless, Brave Search, DuckDuckGo, Parallel, Sprites, Cursor — auto-registered via the inventory plugin system.
• Client-side tools for SDK/API consumers that want to run tools locally.

Ship agents to real channels

Define an App that binds a Harness + Agent to one or more channels: