JetBrains adds Junie /demo, an agent that clicks through your PR and records a video

JetBrains announced Junie /demo on September 18, a new mode in its Junie CLI that turns a pull request into a video demo. You describe what to check, Junie builds and launches the app, drives the UI with a model that supports Computer Use, and hands back an HTML report with screenshots and a video of the run. The team says it has already used /demo on more than 1,500 pull requests internally, produced over 2,100 demo videos, and run more than 1,300 release smoke tests through the same agent.

What a run produces

A /demo run picks the changes to check (branch, session, working tree, or last commit), takes the prompt, and executes it against a live copy of the application. The video opens with a slide describing the scenario, records the interaction, and ends with a summary slide with the observed result. Subtitles are added during post-processing. The HTML report gathers the request, the video, the screenshots, and the pass or fail verdict for each step. JetBrains suggests the format is aimed at reviewers, QA, and teammates who ask how a feature works.

For CI, the same mode is available through junie --demo. The team ships two example GitHub Actions workflows: one that runs on pull requests and comments a link to the artifacts, and one that runs release smoke tests across independent scenarios. Because both use GitHub Artifacts, no separate video hosting is required.

The environment under it

Each run happens in a Docker container, based on Debian Bookworm, with Chromium, Node.js, xterm, an Xvfb virtual desktop, a window manager, xdotool, and ffmpeg. A model with Computer Use drives the app through clicks, keystrokes, and screenshots. Junie prefers the developer's active model when it supports Computer Use, and otherwise falls back through GPT-5.6 SOL, GPT-6 Astra, GPT-5.5, and GPT-5.4, always at High reasoning effort. A repository with several environments can define one Dockerfile per template under .junie/vms/ and let Junie pick the right one from .junie/demo.md.

What it costs

JetBrains published a 22-case internal comparison in the post. Running the full set on GPT-5.6 SOL cost $19.94 across the 22 scenarios, at $1 per AI Credit in JetBrains' subscription conversion, and SOL was the fastest and cheapest of the three models the team tested. Those are JetBrains' own numbers on JetBrains' scenarios: any real project's build steps, prompts and models will change them, and CI runner time is on top.

A demo takes several minutes. The value is that the manual clicking step of a review, the part where somebody launches the app, finds the right screen, and clicks through the flow, moves from the author to the agent and produces evidence a reviewer can watch instead of reproducing themselves. A change without a user-visible surface, such as a backend refactor with no Swagger or admin screen to open, gets less from this than a feature with a UI. The prompt matters as well: JetBrains' own advice is to name the action, the expected state, and one condition that must survive a reload. "Check the feature" leaves too much for the agent to interpret.

Source

• Junie /demo: Turn code changes into video demos, JetBrains Blog, September 18, 2026