ThinkRun began as a way for AI to use the web. Then the web taught it something harder: doing the work is only half the work. The other half is leaving evidence.
The first commit was not a manifesto. It was a machine: sessions, queues, a sandbox, a browser on Fly.io. A simple idea with a difficult edge—let an agent go somewhere, do something, and come back.
The next 323 days turned that edge into an entire product philosophy.
ThinkRun’s earliest shape was pure execution: provision a browser, steer it, collect the result. The browser was remote, the contract was an API, and success meant the task finished.
Real websites do not behave like demos. They redirect, stall, mutate, and fail halfway through. ThinkRun learned to extract the DOM, retry, continue after recoverable failure, stream task plans, and preserve before-and-after evidence.
The browser moved from a remote machine into the user’s real Chrome. A Manifest V3 extension, native host, local bridge, CLI, and MCP server made the boundary disappear: the agent could work where the user already was.
A completed task that nobody can inspect is just a claim. ThinkRun made sessions shareable, embeddable, and machine-readable. One link could carry the replay, screenshots, action trail, console, and context to a human—or another agent.
The product turned inside out. Instead of asking people to reconstruct a bug for an agent, ThinkRun captured the screen, voice, clicks, console, and network together. The agent could read what actually happened, in time.
The last mile became the product: CLI and MCP parity, remote-control grants, a connector gateway, OAuth for remote agents, machine-readable share bundles, search-ready docs, and a recorder-first homepage.
These are not vanity numbers. They are the visible pressure marks left by a product learning in public: code, review, memory, correction, repeat.
Truth is a feature. An earlier timeline draft estimated six figures and called them verified. The repo’s own memory records the correction. Every number above comes from the corrected raw-history recomputation; the mistake remains part of the story because evidence matters most when it proves you wrong.
Agents don’t need more screenshots.
They need a memory.
A screenshot shows a state. A recording shows causality. ThinkRun became the layer that preserves what happened well enough for a person—or an agent—to act on it next.
The development memory is full of hard-won corrections. Underneath the features, four principles became the product’s operating system.
Running, blocked, completed, failed. The system should never make the user guess what the agent is doing.
Every important claim should resolve to replayable evidence: an action, screenshot, event, request, or share artifact.
The best browser is often the one already open. ThinkRun learned to meet users in their real Chrome without losing cloud reach.
Done means shipped, deployed, and verified where users live. Green code in a private branch is only a possibility.
Record the thing that broke. Share what actually happened. Give your agent the context it needs to finish the work.