Open Director is a local-first creative workspace for AI video generations.
All generations are stored directly in your local project folder, just like a coding IDE.
The center of the app is not a cloud dashboard. It is your local filesystem.
Most AI video generation tools are cloud only. This is the wrong paradigm for professional content creators.
Professional content creators use local software such as PR, Davinci, or even CapCut to edit videos and use local file systems to organize their assets.
Local file system is still the most powerful creative vehicle because it allows different software to work together without they knowing each other.
That's why the best code editors and IDEs are still local after the-Cloud have been around for more than 15 years.
That's also why Obsidian is so successful.
Data are just files.
By storing all data locally, there is zero network latency when the user wants to view a video generation. There is no downtime. Users use file systems to organize and manage their assets. Cloud backup can be solved by other tools such as Dropbox or Google Drive or even git to version control their assets.
Plus, because Open Director is local. Your local coding agent such as Claude Code, Codex or OpenClaw can drive it through the MCP interface.
Every tRPC procedure under the open namespace is exposed automatically as an
MCP tool using the same procedure name and Zod-derived input schema. Procedures
outside open remain application-only. Adding a new open procedure therefore
adds the corresponding MCP tool without a second, hand-written registration.
Sharing is also just sharing a folder/zip, instead of a URL that other people have to register an account to have any meaningful access.
Cloud-based AI video generation platforms either can't do it, can't do them all, or can't do them well. (They are all amazing platforms, with no disrespect)
Although, currently, the most powerful models are still closed and can only run in the cloud. But, we should at least have the generation artifacts locally available. Seedance 2 won't be the last powerful AI video generation model. Open Director's architecture makes it easy to integrate any future models, whether open-source or proprietary.
Own your data, own your file, and have fun.
Open Director is intended to run as a trusted local application.
Do not expose it directly to the public internet. The backend can read and write files in the selected project, launch files with the OS default application, store an API key locally, serve project files, and expose MCP and tRPC endpoints.
Image/Video Harness (pattern). An Image/Video Harness is the software assembly that wraps a generative image or video model — a stateless, weakly-reasoning instrument — with the external intelligence, memory, and control it lacks, so that a loosely-specified creative intent reliably becomes a finished asset at minimum cost. Unlike an LLM harness, where the wrapped model is also the brain, an image/video harness is inherently a two-model system: a reasoning model (the agent) drives a generation model (the instrument). A system earns the name "harness" — rather than "wrapper," "client," or "UI" — only when it supplies all four of the following functions, each compensating for a specific deficiency of the generation model:
Context persistence — defeats statelessness. The harness holds the durable specification — base prompt, locked art style, character and asset registries — so intent is declared once and silently re-applied to every call, instead of being re-typed each generation. Prompt synthesis — supplies the missing brain. An external reasoning model translates loose human intent into model-optimal prompts, compensating for the instrument's brittle, shallow text conditioning. This is the function that makes it a two-model system; without it you have a UI, not a harness. Iterative control — closes the expensive loop. The harness runs a generate → judge → regenerate cycle — evaluating output, deciding whether to retry, converging on an acceptable result — because the instrument cannot critique or correct itself and every call costs real money and time. Cumulative memory — improves with use. The harness externalizes learned prompt craft and reusable assets to local, persistent storage — the self-improving skill library — so the system gets measurably better the more it is used, rather than starting cold every session.
The test: a system that provides only some of these is a generation client. A system that provides all four is a harness — and the value it captures scales with the cost of the instrument it wraps, because better prompts and fewer wasted generations matter most precisely when each generation is slow and expensive.
This is an early project, expect breaking changes
No license has been selected yet.
If the goal is to be genuinely open source while discouraging proprietary closed-source forks, this project is not MIT.