Midscene is not sitting still. ToolVitals sees 13 release events in 30 days, 30 GitHub releases in 90 days, a 100 shipping score, and 12,974 GitHub stars. For a developer tool trying to make vision-driven UI automation work across web, desktop, and mobile, that release cadence matters.
The sharp signal is not just frequency. It is direction.
Midscene’s official site positions it as “AI-powered, vision-driven UI automation for every platform.” The home page backs that up with web, PC, Android, iOS, HarmonyOS, Linux, macOS, Windows, and any-interface automation. The GitHub README says the same thing, with natural language automation, JavaScript and YAML workflows, Playwright and Puppeteer integrations, Android and iOS control, MCP, caching, visual reports, playgrounds, and a pure-vision route for UI actions.
That is a bigger bet than browser testing.
Recent releases show the team hardening the product around that bet. v1.7.7 added an assert MCP tool backed by aiAssert, signed Studio mac release builds, Studio packaging work, startup smoke coverage, and report and visualizer fixes. v1.7.6 exposed agent.aiLongPress() and agent.aiClearInput(), merged reports in the CLI, improved dropdown guidance, and cleaned up iOS and Playwright behavior. v1.7.5 added Studio dark mode, a multi-platform playground covering Android, iOS, HarmonyOS, and Computer, site work around MCP and Skills, and several reliability fixes.
This reads like a project moving from “can an agent click the thing” toward “can teams debug, package, replay, and integrate UI automation across real surfaces.” That is the hard part. Vision agents look magical in demos, then get ugly when selectors disappear, viewports shift, popovers move, or mobile state gets weird.
Midscene is building around those ugly parts.
What ToolVitals can and cannot infer
ToolVitals can say Midscene is active. The numbers support that: 95 health score, 100 shipping score, 94 ToolVitals score, 226.8 hot score, 30 releases in 90 days, and 13 release events in 30 days. The repository is public, MIT licensed, and the project’s site says it is free and open source.
ToolVitals cannot say Midscene’s automations are reliable in your app. It does not run your test suite. It does not measure task success rate, model cost, latency, false positives, user satisfaction, revenue, or whether aiAssert catches the failures you care about.
The data confidence is 84, so the maintenance signal is solid but not omniscient. Commits in the last 30 days and active contributor counts are null in this payload, so this post should not infer team size or commit velocity from ToolVitals.
How it compares
Midscene is smaller than the largest adjacent developer tools in raw GitHub attention. LangChain has 136,337 stars and 32 release events in 30 days. Gemini CLI has 103,730 stars and 17 release events in 30 days. Midscene has 12,974 stars and 13 release events in 30 days.
That gap matters, but it cuts both ways. Midscene is not a broad AI developer platform like LangChain, and it is not a general CLI surface like Gemini CLI. It is narrower: vision-driven UI automation. For a focused tool, matching the same 100 shipping score while pushing 30 releases in 90 days is the useful comparison.
Against ToolJet, Midscene is also smaller by stars, 12,974 versus 37,888, and slightly behind on 30-day release events, 13 versus 15. Both show a 100 shipping score. The difference is category pressure: ToolJet competes in app building, while Midscene is trying to make UI control itself programmable through vision and agent APIs.
Recommendation
If your team is experimenting with AI-driven UI automation beyond ordinary browser selectors, evaluate Midscene now. The current signal is strongest for teams that need web plus mobile or desktop coverage, want MCP integration, and care about debugging artifacts like reports, playgrounds, and replay.
Do not adopt it only because the metrics are hot. Pilot it on one brittle workflow, measure task success rate, cost, and failure modes, then decide. The ToolVitals signal says Midscene is alive and shipping quickly. Your own tests have to prove whether it is dependable enough for production automation.