A story with receipts · 2025—2026

The browserlearned to speakagent.

This is the story of 323 days spent closing the distance between what happened in a browser—and what an AI agent can understand, verify, and fix.

Scroll to run
The original problem

A screenshot shows the scene.
A recording remembers the crime.

Agents could write code. Browsers held the evidence. The missing piece was a reliable witness that both could understand.

ThinkRun began as browser automation: give an agent a destination, let it navigate, click, fill, and extract. The first phase was small—38 commits—but it proved the central idea. A browser could become an instrument, not a black box.

app.example.com / checkout
Execution log
00:01NAVIGATEOpen checkout
00:04THINKFind primary action
00:06CLICKContinue
00:08CAPTUREKeep the evidence
38commits in genesis
#1first merged PR
1browser, made legible

The prototype became an engine. Authentication. Resilient tasks. Video playback. Then the decisive turn: a Chrome extension and local bridge let agents work in the browser people already use—the one with their tabs, their sessions, their reality.

The week the machine acceleratedCOMMITS / ACTIVE WEEK · VERIFIED
AUG 2025MAR 16–22 · 155 COMMITSMAR 2026
Before

A fresh browser.

The agent arrived logged out, contextless, and separated from the person asking for help.

cloud browser
+ clean profile
− your session
− your tabs
The local turn

Your real Chrome.

Scoped tab ownership, attach receipts, policy-checked commands, and an always-available kill switch.

tab attached
session inherited
policy on
human in control

Control was only half the story. ThinkRun learned to capture screen, voice, clicks, console, and network as one time-synced artifact—then share it as a link, Markdown, JSON, or agent context. The browser stopped being somewhere an agent went. It became something the agent could recall.

thinkrun.ai / s / the-whole-story
One recording · many readers

A bug report
your agent can watch.

Screenwhat changed
Voicewhat you meant
Tracewhat actually failed
The part most launch stories omit

The agent
was wrong.

The first timeline confidently labeled estimates as verified. It even invented a plausible explanation for arithmetic that did not reconcile. Independent recomputation caught it.

That failure belongs in the story. ThinkRun was not built by treating agent output as truth. It was built by demanding evidence: recordings, receipts, review gates, source trails, and numbers that add up.

Total commits1,5701,571
PRD threads5962
Phase 3 PRs~350147
Corrected against raw git on July 9, 2026. No lifetime cost or token figure is shown because the repository does not contain complete metering for the 323-day history.

What the
receipts say.

Every number below comes from the repository’s independently refreshed timeline, recomputed from raw history on July 9, 2026.

0
commits across six development phases
audited 2026-07-09 timeline snapshot
0
unique merged pull requests
deduplicated PR numbers in snapshot subjects
0
days from first commit to verified timeline
2025-08-20 → 2026-07-09
0
commits in the peak week
2026-W12 · Mar 16–22
0
memory pages and indexed entries consulted
60 threads + 64 daily logs + 140 entries
The run continues

The browser is
no longer mute.

It can show what happened, preserve why it mattered, and hand the whole story to the agent ready to do something about it.

Meet ThinkRun