Agenvas
Self-hosted AI creation canvas
About Agenvas
Agenvas is an open-source, self-hosted AI creation canvas. Text, images, video, audio and agents all live as cards on one infinite canvas, where they can reference each other, be regenerated, versioned and arranged into a complete creative flow. It is MIT licensed, runs on your own machine or server through Docker Compose, and is built for a single administrator who wants the feel of a commercial AI canvas without paying for an expensive subscription. You connect your own model providers, so the choice of which models to use, and how much to spend on them, stays with you.
The project is written by an independent developer and published on GitHub at grayrepo-byte/Agenvas, with a Java 21 and Spring Boot backend (Spring AI for the model and agent runtime), a Vite, React and TypeScript frontend that uses React Flow for the canvas, and PostgreSQL for storage. The repository ships tagged releases, a product spec, a development checklist, backup and restore notes, a Windows install guide and a security policy, and the interface is available in Chinese, English, Russian and Japanese. It is young: the repository was created in September 2026 and the README describes the current build as a development version for single-admin self-hosting, with no guarantee of data compatibility between future versions. There are no paid plans, no membership tiers and no usage threshold; the only costs are your own hosting and whatever your chosen model providers charge.
You install Agenvas with a single command. On macOS or Linux the quick start creates an agenvas directory, downloads the docker-compose.yml file from the repository and runs docker compose up, which pulls the images and waits for the web, API and PostgreSQL services to become healthy. Windows users get equivalent CMD and PowerShell one-liners. Starting it needs no Git, no local build toolchain, no GPU, no .env file and no model keys, only Docker Engine or Docker Desktop with Docker Compose v2. The database password and the credential encryption key are generated automatically on first start.
Next you open the setup page on port 8088 and create the admin account. Setup completion is recorded permanently in the database, so restarts never reopen it, and the project recommends finishing setup on a controlled network before exposing the service to the internet.
Then you configure models. Under model configuration you add the endpoint, model ID and API key for any OpenAI-compatible provider, which powers text and the agent. Under media configuration you create connections, publish image, video and audio capabilities and set default models. Finally you create a project, add cards or upload media from the canvas context menu, pick a model, write a prompt, add reference inputs and run the card. Results can be previewed, selected or regenerated, and every card keeps its version history.
- •
Canvas Creation - Every asset is a card on an infinite canvas. You can drag, box-select, connect and auto-arrange cards, upload local media, and see at a glance how prompts, images, video, audio and agent cards relate to each other inside a project.
- •
Multimodal Generation - Text, image, video and audio run per card, with queuing, cancellation, regeneration and full version history. Text works against any OpenAI-compatible endpoint with a custom model ID. The site lists GPT Image, Google Nano Banana, Volcano Ark Seedance, Seed Audio, ComfyUI workflows and RunningHub V2 among the supported adapters and workflow engines, and video results in an unknown state must be retried explicitly so an unfinished job is never mistaken for a finished one.
- •
Agent Collaboration with Approval - The agent reads the context bound to it, edits text and places cards on the canvas. When it wants to generate media it submits proposals, and those only run after you approve them, so you stay in control of what gets generated. Administrators are asked to run a tool-calling diagnostic to confirm a model is ready for agent use.
- •
Asset Reuse - Project resources, a personal asset library, prompt templates, media styles and mixed reference inputs are saved once and reused across cards and projects, so you are not re-uploading the same references or rewriting the same prompts.
- •
Image Processing - Images can be annotated, cropped, rotated, mirrored, upscaled and given depth extraction directly on the canvas, plus AI edits wherever the configured capability supports them.
- •
Self-Hosted Administration - The admin controls model configuration, storage (local disk or OSS, COS and S3 object storage), task and call records, and project manifest export. API keys are encrypted at rest and kept out of logs, and the database password and master key are generated and persisted for you.
Agenvas is built for individual creators and tinkerers who like the idea of a node or card based AI canvas but do not want to pay a monthly membership for one, or be tied to the most expensive models. If you already have API access to an OpenAI-compatible model provider, or run ComfyUI workflows yourself, Agenvas gives you one place to combine text, images, video and audio into a project and keep every version.
It suits people producing short video and multimedia pieces, such as the short drama project shown in the product preview, where a prompt spawns text and images, an agent proposes new shots and a voice-over, and approved proposals become audio and video cards. Developers and self-hosters who want to keep their prompts, media and keys on their own server will appreciate the Docker Compose setup, encrypted credentials and choice of local or S3-compatible storage. Because the code is MIT licensed and the stack is Java, Spring Boot, React and PostgreSQL, it is also a readable starting point for anyone who wants to study or extend an AI canvas. It is less suited to teams that need multi-user accounts or a stable, long-term data format today, since the project describes itself as a single-admin development build.
