Mirage, the company that created the Captions app, launched Tesseract on September 22, 2026 — a free creative suite for AI agents. Based on briefs and references, the agent edits video, assembles motion graphics and design, working directly with an editable project rather than one-off clip generation. The native engine runs on macOS, Windows, and Linux, local previews and rendering are free, and no Adobe subscription is required.

image
image

What happened

On September 22, 2026, Mirage introduced Tesseract as a creative suite controlled by an LLM agent. Based on briefs and references, the agent operates a full project model: it edits text, shapes, masks, keyframes, adjustment layers, track mattes, timing, and audio, so changing one element does not require regenerating the entire video. Ready-made agents such as ChatGPT, Claude, Muse, Grok Bot, and Instinct can be connected for control, and monetization has been moved up the stack: the engine itself is free, and only the agent subscription is paid for. For custom agents, Mirage has made CLI builds and agent skills publicly available in the mirage-hq/Tesseract repository, with Python code updated on October 1, 2026.

Context

Tesseract is not built on a new generative model, but on a mature editing engine: according to Mirage, the same engine powers the Captions app, through which millions of videos have already been assembled. The launch is interface-driven rather than model-driven: the engine with a full project model is given to agents for free, and all the "intelligence" remains with the connected LLM. In effect, agentic work is being moved from the paradigm of one-off clip generation to an editable project model, where each command is fixed to a specific layer or keyframe. The difference from familiar text-to-video services is that the result remains a live project that is refined to the final version through iterations with local previews, rather than being discarded after a single generation.

Why this matters for the industry

For builder teams, the signal is direct: the editor layer is being commoditized, and orchestrating a free engine is more cost-effective than developing your own editing core. A startup can build an MVP video production on top of open CLI builds and agent skills without investing in NLE engineering from scratch. Point editing of a layer or keyframe is cheaper than regenerating a clip in its entirety, so the competitive advantage shifts to the quality of orchestration and agent selection, rather than the engine itself. Over the next six months, the maturity of converters and the activity of the custom agent community will be decisive: if open skills build an ecosystem, the editable project model could de facto become the standard integration point between LLM providers and project files. Over the next couple of years, stack diversification is likely, where generative models produce footage, agentic editors assemble it into a project and finish the graphics and timing, and the main risk for Mirage is that Adobe could natively embed an agentic layer into its already familiar and compatible ecosystem.

Why this matters for users

You can try it for free right now: you provide a brief, references, a brand kit, and your own assets, and in return you get an editable project with local previews and point edits up to the final render at up to 4K 60 fps resolution. The entry point is the mirage.app/tesseract page with instructions and a ready-made setup prompt, and CLI builds and skills are in github.com/mirage-hq/Tesseract. For now, the most reliable tasks are those that live entirely within the Tesseract ecosystem: editing your own assets, graphics, titles, and iterative edits, including commands like "move the title two seconds earlier." However, Tesseract does not generate footage or avatars, and all source material must be your own.

What is still unknown / limitations

Trust in Mirage's claim of "millions of videos" in Captions should be calibrated: this is a vendor statement on its own product page, with no independent data. The accuracy of agentic manipulation of keyframes and timing has not been verified by benchmarks, error rates and comparisons with manual editing are unknown, and the final quality of edits is entirely determined by the connected agent. Conversion to/from Premiere Pro and After Effects is partial, AE export remains experimental, so migrating existing Adobe projects needs to be tested on your own files. The dynamics of converter maturity can be tracked by activity in the mirage-hq/Tesseract repository.

Sources

Author

Look at AI, editorial team