Who OpenMAIC is for#
Corporate trainers building AI role-play simulations
OpenMAIC generates interactive training scenarios from uploaded scripts, call recordings, or written briefs. AI agents play customer, manager, or peer roles in the simulated conversation and trainees can respond and get real-time feedback. This covers the core use case of paid platforms like Second Nature, without the per-seat SaaS cost.
Skip if:
If your training requires certified human coaches, structured assessment reporting, or LMS integration out of the box, OpenMAIC does not ship those features and would need custom development to meet compliance requirements.
Educators creating interactive lessons from course materials
Upload lecture notes, syllabi, PDFs, or audio recordings and OpenMAIC generates a full lesson with slides, quizzes, and PBL activities. The agent workbench lets you iterate on the curriculum interactively before finalizing, and the .pptx and .html exports make content portable.
Skip if:
If you need a learning management system with grade tracking, student accounts, and assignment workflows, OpenMAIC is a lesson generator rather than a full LMS and the missing features would require building on top of it.
Developers building AI-powered learning products
The @openmaic/* SDK (DSL, renderer, importer) is published to npm, so teams can embed or customize the classroom engine in their own products. Per-stage model routing lets you assign different LLMs to different generation tasks, and the pluggable provider architecture means you control cost and capability tradeoffs.
Skip if:
If you need a managed API with SLA guarantees, a support contract, or SCORM compliance out of the box, OpenMAIC is an open source engine rather than a managed service and those requirements would need additional work.
Teams self-hosting AI training with data privacy requirements
Running OpenMAIC on your own infrastructure means training content, conversation history, and model calls stay on your servers or within your cloud account. Bedrock, Ollama, and Lemonade are all viable LLM backends that keep inference in-network. The Postgres persistence stack and incremental save system support multi-session operation.
Skip if:
If you process personal data subject to strict regulatory regimes (HIPAA, EU AI Act high-risk classification), verify the model providers you configure meet those compliance requirements. OpenMAIC itself is infrastructure-agnostic.
The problem it solves#
Immersive AI training simulations have been locked behind enterprise SaaS pricing and proprietary infrastructure. Platforms built for role-play practice and interactive learning charge per-seat fees, own your training content, and give you no access to the underlying AI models or curriculum generation pipeline. A team running hundreds of training sessions or a course creator building immersive lessons pays continuously for something they cannot modify, export, or host themselves.
The deeper problem is inflexibility. Paid training platforms ship fixed AI personas and fixed content templates. Swapping to a newer or cheaper LLM is not possible, connecting your own document library is not supported, and adding a custom scene type requires vendor permission. When pricing changes or a feature disappears, there is no fallback. Teams that need immersive AI learning at scale, with their own data and their own models, have had no credible open source path.
How it solves it#
One-click lesson generation from any topic or document
Describe a subject or attach a file and OpenMAIC generates a complete lesson: slides, quizzes, project-based learning activities, and interactive simulations. The generation pipeline runs against your chosen LLM, so output quality and cost scale with the model you configure.
Multi-agent classroom with AI teachers and peers
A network of AI agents plays teacher, classmate, and facilitator roles simultaneously. Each agent speaks via text-to-speech, draws on a shared whiteboard, and responds to your input in real time. The whiteboard supports diagrams and formulas, not just text.
Rich scene types: slides, quizzes, simulations, and PBL
Lessons are composed of typed scenes: presentation slides (exportable to .pptx), quizzes with persistent state, interactive HTML simulations with 3D visualization and online programming, and project-based learning (PBL) activities with a structured classroom UI.
Provider-neutral LLM and media stack
Connects to OpenAI, Anthropic, Google Gemini, Amazon Bedrock, Azure OpenAI, DeepSeek, Qwen, Kimi, MiniMax, Grok, Ollama, Lemonade, and 10+ others. Each capability (language model, TTS, ASR, image generation) is independently configurable, so you can mix providers per role.
Agent workbench with durable, steerable sessions
The v1.0.0 Pro workbench is a chat-first interface where you plan the curriculum with an agent, upload session materials (documents, audio, video, or web search), and steer generation interactively. Sessions are server-backed and survive restarts; you can cancel, resume, or redirect at any point.
Local and offline AI with Lemonade and FunASR
Run the full classroom stack without external API keys. Lemonade provides a local OpenAI-compatible endpoint for LLMs, image generation, TTS, and ASR. FunASR adds local speech recognition (SenseVoiceSmall, Paraformer, Fun-ASR-Nano) for transcribing uploaded audio and video materials.
Strengths and trade-offs#
Strengths
- MIT licensed with no usage limits on the self-hosted pathThe entire codebase is MIT licensed, so you can run it commercially, fork it, and modify it without restriction. Unlike hosted training platforms that charge per session or per seat, a self-hosted deployment has no per-use cost beyond your LLM API spend.
- Actively maintained with research backingPublished in the Journal of Computer Science and Technology (JCST, 2026) by Tsinghua University researchers, with over 22,000 GitHub stars and a release cadence that added local ASR, video export, a Pro editor, and an agent workbench across five major versions between March and August 2026.
- Provider-neutral: bring any LLM, TTS, or image backendNo single LLM is required. The configuration accepts 15+ providers for language models alone, plus separate providers for TTS, ASR, image generation, and video. Switching from OpenAI to Gemini or adding a local Ollama instance requires only a change to the environment file.
- Full local deployment for air-gapped and privacy-first teamsLemonade and FunASR together cover all media modalities locally: LLM inference, TTS, ASR, and image generation with no external API calls. Teams with data privacy requirements or air-gapped environments can run the entire pipeline on their own hardware.
Trade-offs
- -Requires Node.js 20+, pnpm, and at least one LLM provider configuredThe development path needs Node.js 20 or newer and pnpm 10+. The Docker Compose path simplifies this, but you still need to configure at least one LLM provider API key (or a running Lemonade or Ollama server) before the classroom generator works. This is not a zero-config install.
- -207 open issues reflects a fast-moving, young projectOpenMAIC reached v1.0.0 in August 2026 after launching in March 2026. The issue count (207 open as of late August 2026) reflects rapid feature development rather than neglect, but teams evaluating it for production use should expect rough edges and breaking changes between versions.
- -Local audio/video extraction requires ffmpeg and a separate ASR serverTranscribing uploaded audio or video files locally requires system ffmpeg on PATH plus a configured ASR provider (FunASR, Lemonade, or OpenAI). The app skips local extraction gracefully when ffmpeg is missing, but teams wanting fully offline media processing face a non-trivial setup.
OpenMAIC vs alternatives#
OpenMAIC vs Mursion
Mursion delivers avatar-based immersive simulation training for workplace skills (leadership, sales, customer service) via a proprietary SaaS platform. It combines pre-built AI avatars with human simulation specialists who can intervene during sessions. OpenMAIC generates multi-agent classrooms from arbitrary topics using any LLM you configure, without pre-built personas or human involvement.
| Feature | OpenMAIC | Mursion |
|---|---|---|
| License | MIT (self-hosted) | Proprietary |
| Self-hosting | Yes | No |
| LLM provider | Configurable (15+) | Proprietary AI |
| Pricing | Free self-hosted | Enterprise (quoted) |
| Content types | Slides, quizzes, PBL, simulations | Avatar role-plays |
| Human facilitation | No | Optional (simulation specialists) |
OpenMAIC is the better choice when you need to generate classroom content from your own documents and run the simulation pipeline on your own infrastructure. It handles broader content types (not just role-play) and lets you swap LLMs as the model landscape evolves. Mursion is worth considering when your training program specifically requires certified human simulation specialists for high-stakes scenarios (clinical, leadership development) where a managed, observable session is necessary.
OpenMAIC vs Second Nature
Second Nature is a conversational AI sales training platform where AI personas practice pitches and handle objections alongside sales reps. It is purpose-built for sales enablement with built-in pitch libraries, objection banks, and CRM-connected reporting. OpenMAIC is a general-purpose AI classroom engine that can simulate similar conversational training scenarios but requires you to configure the content and personas.
| Feature | OpenMAIC | Second Nature |
|---|---|---|
| License | MIT (self-hosted) | Proprietary |
| Self-hosting | Yes | No |
| Use case focus | General AI classroom | Sales training |
| LLM provider | Configurable | Proprietary |
| CRM integration | None (native) | Yes |
| Pricing | Free self-hosted | Per-seat SaaS |
OpenMAIC wins for teams that want to build AI training scenarios across multiple domains (sales, customer service, onboarding, technical training) from their own content, host the pipeline privately, and avoid per-seat fees. Second Nature is the better pick when you need a purpose-built sales training tool with pitch library management, CRM integration, and onboarding support out of the box, and the per-seat cost fits your budget.
Install and self-host#
Self-hosting uses Docker Compose after cloning the repo and configuring at least one LLM provider API key in .env.local.
```bash
git clone https://github.com/THU-MAIC/OpenMAIC.git
cd OpenMAIC
cp .env.example .env.local
docker compose up --build
```What it's built on#
- Languages
- JavaScriptTypeScript
- Frameworks
- Next.jsReact
- Databases
- PostgreSQL
- Infrastructure
- AWS
- Tooling
- Rollup
FAQ#
Is OpenMAIC free to use?
Yes. OpenMAIC is MIT licensed and free to self-host on your own infrastructure, with no per-session fees or user limits. You pay only for any external LLM, TTS, or ASR API calls your configuration makes. A fully local setup using Lemonade and FunASR eliminates external API costs entirely.
What LLM providers does OpenMAIC support?
OpenMAIC supports OpenAI, Anthropic, Google Gemini, Amazon Bedrock, Azure OpenAI, DeepSeek, Qwen, Kimi, MiniMax, Grok (xAI), GLM (Zhipu), Xiaomi MiMo, Tencent Hunyuan, Doubao, OpenRouter, Ollama, and Lemonade, plus any OpenAI-compatible API. Provider configuration is per capability: you can use different backends for language generation, TTS, ASR, image generation, and video.
How is OpenMAIC different from Mursion or Second Nature?
Mursion and Second Nature are paid commercial platforms for AI-driven training simulations, hosted on proprietary infrastructure with no self-hosting option. OpenMAIC is open source and MIT licensed: you host it on your own servers, bring your own LLM, and own your training data. The tradeoff is that OpenMAIC requires technical setup, while Mursion and Second Nature are managed services with onboarding support.
Can I use OpenMAIC without any external API keys?
Yes. Running Lemonade locally gives you a fully offline LLM, TTS, ASR, and image generation stack compatible with OpenMAIC's provider system. FunASR adds a local speech recognition server for audio and video materials. Expect higher hardware requirements compared to cloud API-based setups.
How do I deploy OpenMAIC in production?
The recommended paths are Docker Compose (clone the repo, configure .env.local with your API keys, then run docker compose up --build) and Vercel one-click deploy for teams that want a managed frontend. For shared deployments, set ACCESS_CODE in .env.local to gate access with a site-level password. The Postgres persistence stack is required for durable multi-session agent workbench use.
Similar open-source tools#
garden-skills
Production-ready agent skills for Claude Code, Cursor, and Codex
browser-use
Python library giving any LLM full browser control, MIT licensed
free-claude-code
Route 9 AI coding agents through 1.3B+ free monthly tokens
OpenCompany
Self-hosted AI agent canvas for every business function
codex
OpenAI's terminal coding agent, Apache-2.0 licensed
thoughtdag
Make LLM context visible, editable, and graph-structured