This is the recorded history of a tool that began by giving agents a browser—and ended up giving them eyes, evidence, and a memory.
“It looks wrong” is not a bug report. A screenshot is not a timeline. A transcript is not the browser. ThinkRun grew around one stubborn idea: let an agent see the exact moment humans see—and preserve the evidence all the way to the fix.
Every chapter added a missing sense. The browser became controllable, then local, then shareable, then recordable, then legible to any agent.
On August 20, 2025, the first commit opened the door: a browser automation service, session management, queue integration, and deployment. It was not yet a story. It was a primitive—the ability to act.
Authentication changed. Tasks learned to retry. Sessions gained video, artifacts, and before-and-after screenshots. The agent stopped being a demo and became a system that could survive failure.
The cloud was useful. Your already-signed-in Chrome was transformational. A Manifest V3 extension, native host, and local bridge let agents work where people already worked—inside the tabs they deliberately attached.
A session could now leave the person who recorded it without losing its meaning: embeddable video, oEmbed, reliable uploads, structured feedback, and MCP workflows. One link carried the scene of the bug.
Upload any video. Record in the page. Generate playback derivatives. Bring the session into an Activity Hub. The product crossed the line from controlling browsers to translating lived experience into agent-ready context.
CLI and MCP parity, remote connectors, SEO, onboarding, pricing, and production hardening turned the internal machine outward. ThinkRun was no longer only a browser tool. It was a shared language between a person, a screen, and an agent.
Verified from raw git history and the repo memory wiki through July 9, 2026. The six phase commit totals—38 + 279 + 520 + 280 + 306 + 148—sum exactly to 1,571. The story drew from 264 memory pages.
The memory did not just remember the work. It caught the story lying.
An early timeline confidently labeled stale estimates as verified. It even invented a plausible explanation for the mismatch. The wiki’s receipts forced a recomputation. Six phase totals changed; the fake discrepancy disappeared.
So this page keeps the scar visible.ThinkRun’s deeper story is not “AI builds fast.” It is that evidence has to survive the telling. The same principle that makes a bug report trustworthy must make a company story trustworthy, too.
Point at what you mean. Keep the moment intact. Hand your agent the whole picture—and let the work begin from evidence, not approximation.