An origin story · verified through July 9, 2026

We taught
software to see.

ThinkRun began with a stubborn idea: when something breaks in a browser, an agent should receive the evidence—not a paraphrase of the evidence.

thinkrun.ai / a browser remembers
Session replay Ready for agent
  • SCREENThe page, as it happened.
  • THINKThe intent, not just the pixels.
  • ACTIONA path from evidence to a fix.
Scroll to replay

The premise

A browser is not a page.
It is a place where intent leaves evidence.

The first commit landed on August 20, 2025. The premise fit in one sentence. The work did not.

The build log

The browser had to become legible.

Not a bot that clicks. A system that can perceive, act, recover, and leave enough evidence for the next person—or the next agent—to understand what happened.

01 / GENESIS
AUG 20 — NOV 30, 2025

A window opens.

Initial browser automation turned navigation, screenshots, and extraction into the first primitive verbs of a new sense.

Git receiptGenesis
38commits
02 / CORE ENGINE
DEC 1, 2025 — JAN 31, 2026

Seeing becomes doing.

Identity, task resilience, and playback made browser work durable enough to survive more than a happy-path click.

Git receiptCore engine
279commits
03 / LOCAL
FEB 1 — MAR 31, 2026

The agent enters your Chrome.

The extension and local bridge changed the boundary: the agent could work in the browser session you already trusted.

Git receiptExtension + local
520commits
04 / SHARING
APR 1 — MAY 15, 2026

A session becomes a language.

Share links, oEmbed, MCP parity, and hybrid upload let a replay carry its context between people and tools.

Git receiptSharing + MCP
280commits
05 / RECORDING
MAY 16 — JUN 17, 2026

Show, don’t prompt.

Screen, voice, clicks, console, and network became a synchronized handoff—so “watch this” could finally be complete.

Git receiptRecording pipeline
306commits
06 / DISTRIBUTION
JUN 18 — JUL 9, 2026

The tool learns to travel.

SEO, connectors, remote control, and a pricing source of truth moved ThinkRun from a product into a place agents could enter from anywhere.

Git receiptDistribution
148commits

The product in miniature

Stop narrating the bug. Replay it.

Choose an evidence layer. The story changes because the artifact changes.

Inspect evidence

Screen replay

The page tells the first half.

A replay gives an agent the moment as you saw it: the layout, the cursor, the sequence. It replaces “it looks wrong” with an inspectable starting point.

SCREENSHOT viewport captured
ready
CLICK interaction recorded
ready
REPLAY sequence preserved
ready

The receipts

No mythology. Just the work.

All figures are from the independently recomputed July 9, 2026 snapshot: raw Git history first, then the memory wiki for context.

0
commits in the verified snapshot
0
unique merged PRs
0
days from first commit to snapshot
0
PRD thread pages
0
commits in the peak week · Mar 16–22, 2026

The second source of truth

The product kept a memory.

Git says what changed. The memory explains why: a CORS allowlist drift, a dropped pixel, a broken install command, an auth outage, and the decisions that make the same failure less likely next time.

The interesting part is not that ThinkRun has history. It is that the history can answer back.

ThinkRun
memory
Local bridge
recovery
Recording
pipeline
R2 CORS
drift
Auth signup
outage
Pricing
source of truth

The most ThinkRun part

Even the timeline had to pass review.

An earlier story artifact called several estimates “verified,” including an invented explanation for numbers that did not add up. Independent recomputation caught it.

That correction belongs in the story. A product built for observable work has to be observable about its own claims.

Independent verificationJul 9, 2026
Total commits1,5701,571
PRD threads5962
Phase one PRs~254
Phase three PRs~350147
Strict phase total1,571 ✓
Rule learned: verify the artifact, not the claim.