Back to catalog
Agenvas
MITAgenvas is a self-hosted MIT AI creative canvas integrating ComfyUI, RunningHub, Kling, MiniMax H3, OpenAI APIs, and MCP agents for multimodal media.
Open-source alternative to — compare it against:
Overview
Agenvas pairs a React Flow workspace with a Java 21, Spring Boot, and PostgreSQL backend to orchestrate text, image, video, and audio generation in one self-hosted stack. Rather than shipping a proprietary model, it connects OpenAI-compatible APIs, ComfyUI workflows, and media endpoints on an infinite node board. It is built for creators who want to wire together multiple AI providers without recurring SaaS fees.
Why Agenvas instead of Runway Gen-2
Suites like Runway Gen-2, Midjourney, Adobe Firefly, and Canva AI Image Generator tie their interfaces to proprietary cloud inference and credit tiers. That lock-in makes it awkward to mix providers—such as drafting prompts with a local LLM, generating frames in ComfyUI, and rendering clips via Kling or Veo—without shuttling files across tabs. Agenvas decouples the canvas from inference: admins register API keys once, publish capabilities, and route cards to any local or cloud backend.
Agenvas imports ComfyUI graphs and RunningHub V2 workflows as reusable nodes alongside adapters for GPT Image, Seedance, MiniMax H3, Kling, Wan, Seedream 5.0, Veo, and Seed Audio. It also exposes a Model Context Protocol (MCP) server so agents can drive the board. Unlike Runway Gen-2 or Midjourney, Agenvas includes no hosted GPU compute or weights, so you supply and pay for every connected API or server.
What it does well
- Multi-provider adapters: Connects OpenAI-compatible endpoints, Kling, MiniMax H3, Seedance, Wan, Seedream 5.0, Veo, and Seed Audio, plus ComfyUI graphs and RunningHub V2 apps.
- Built-in MCP server: Exposes 13 tools—including voiceover synthesis and media generation—over token-authenticated MCP endpoints for automated pipelines.
- Director agent with cost gates: Embeds a Spring AI assistant with context compaction and 24-hour recovery that binds versions to Skill inputs and queues proposals for approval.
- Character consistency and local tools: Tracks character references, voice samples, and shot drafts, paired with FFmpeg and JNI depth-map extraction and cropping.
- Pluggable storage and relay: Persists assets to local volumes or S3 stores (Alibaba Cloud OSS and Tencent Cloud COS) and relays local files to video APIs.
- Starts as a Compose stack (, , PostgreSQL 17) that generates a database password and Base64 key on first boot, exposing port .
