ARP — Agent Runtime Platform

The open-source runtime that powers AI employees.
Connect any model, any tool, any enterprise system — on your own infrastructure.

---

What is ARP?

ARP is an enterprise-grade, self-hosted agent runtime. It is where AI employees actually run: multi-provider LLM access, no-code agents, MCP tool servers, generative UI, sandboxed code execution, and enterprise-grade access control — deployed on your own servers.

ARP is one component of the Open Insight platform:

Component	Role
openinsight	One-command deployment & release entry
ARP (this repo)	Agent runtime — where AI employees run
one-pi	Enterprise agent platform — expert agents & orchestration

---

Why ARP?

• Any model — OpenAI, Anthropic, Google, Azure, Bedrock, Vertex AI, and any OpenAI-compatible endpoint (GLM, Qwen, DeepSeek, Kimi, Ollama, …).
• Any tool — MCP tool servers, sandboxed code interpreter, web search, image generation.
• Any skill — reusable skills that agents invoke at runtime, shared with ONE-PI: build once, every agent on the platform can call it.
• Self-hosted — your infrastructure, your data, no vendor lock-in.
• Built for AI employees — integrates with ONE-PI (agent platform) and DMP (enterprise data & governance) to ground agents in your business.

---

Screenshots

---

Architecture

Layer	Path	What it does
Web client	/client	React front-end — chat, agents, artifacts, presets
API server	/api, /packages/api	REST & streaming API, auth, integrations
Shared packages	/packages/*	data-provider, data-schemas, shared utils
Configuration	librechat.yaml	Endpoints, MCP servers, interface options
Deployment	docker-compose.yml, deploy-compose.yml, helm/	Docker, Compose, Helm

---

Features

• Multi-provider chat: OpenAI, Anthropic, Google, Azure, AWS Bedrock, Vertex AI, and any OpenAI-compatible endpoint (Ollama, DeepSeek, GLM, Qwen, Kimi, …).
• Agents & MCP: no-code custom agents, Model Context Protocol tool servers, agent marketplace, and remote agent sharing.
• Agent collaboration: @mention any agent into a conversation — agents cooperate in one thread; semantic routing to the right expert is resolved by ONE-PI.
• Generative UI / Artifacts: React, HTML, and Mermaid diagrams rendered in chat.
• Code Interpreter API: sandboxed Python, Node, Go, C/C++, Java, Rust, and more.
• Multimodal & files: image understanding, file chat, SharePoint picker.
• Multilingual UI with reasoning-model support.
• Multi-user, secure access: OAuth2, OpenID, SAML, LDAP, email login, plus the automatic SSO layer.
• Enterprise layer: DMP integration (user lookup, agent context, MCP tool bridge), UI watermarks, branded Sandpack artifacts, custom endpoint presets.
• Resumable streams, conversation search, presets, message branching, import/export.

---

Quick Start

The recommended entry point for the whole platform is openinsight — one command brings up ARP, ONE-PI, and the data layer as a working stack. The steps below run ARP standalone, for development or custom deployments.

Prerequisites

• Node.js v20.19.0+ / ^22.12.0 / >= 23.0.0
• MongoDB
• (Optional) Meilisearch, Redis

1. Configure environment

cp .env.example .env

Open .env and fill in at least:

• MONGO_URI
• JWT_SECRET, JWT_REFRESH_SECRET (generate with node -e "console.log(require('crypto').randomBytes(48).toString('hex'))")
• CREDS_KEY, CREDS_IV (see .env.example for generation commands)
• Any model-provider keys you plan to use

ARP-specific variables (DMP, AUTO_SSO, WATERMARK, SANDPACK, CSP, PI, …) are documented in the OpenInsight Extensions section at the bottom of .env.example.

2. Configure endpoints (optional)

cp librechat.example.yaml librechat.yaml

Edit librechat.yaml to define custom endpoints, MCP servers, and interface options. See librechat.example.yaml for all available settings.

3. Install & run

npm run smart-reinstall   # install deps + build workspaces
npm run backend           # start the API server on :3080

In a second terminal:

npm run frontend:dev      # Vite dev server on :3090

Or run everything with Docker:

docker compose up -d

---

Development

Command	Purpose
npm run backend:dev	Start backend with file watching
npm run frontend:dev	Start frontend dev server (HMR)
npm run build	Build all workspaces via Turborepo
npm run build:data-provider	Rebuild packages/data-provider after edits
npm run lint / npm run lint:fix	ESLint
npm run format	Prettier
npm run test:all	Run all workspace tests

Monorepo layout: /api (legacy JS backend), /packages/api (new TS backend), /packages/data-schemas, /packages/data-provider (shared), /client (frontend), /packages/client (shared frontend utils).

---

Built on LibreChat

ℹ️ Technical note: ARP is built on a fork of LibreChat. Upstream features and general usage are inherited from it.

ARP would not exist without the outstanding work of Danny Avila and the LibreChat contributors.

---

License

Released under the Apache-2.0 License.

Portions of this project derive from LibreChat (MIT, © 2023 Danny Avila and LibreChat contributors).