A development story · 2025—2026

The browser was never
the hard part.

The hard part was teaching an agent what a moment means—and making that moment useful to the next human or agent who sees it.

Open the flight recorder
The web was built for eyes, hands, and context. Agents arrived with none of them.

The flight recorder

Six versions of seeing.

Every phase solved another layer of the same problem: perceive, act, recover, remember—and leave a trail somebody else can use.

01 / GENESIS

A window opens.

Navigation, screenshots, extraction: the primitive verbs of a new sense. 38 commits turned an idea into browser automation.

AUG 20 — NOV 30, 2025
thinkrun://first-light
$ thinkrun navigate example.com
→ browser provisioned
→ page loaded
✓ screenshot captured

The agent can see a page.
It does not yet understand failure.
02 / CORE ENGINE

Seeing becomes doing.

Tasks needed identity, resilience, and playback—not just clicks. ClearAuth, obstacle recovery, and video made browser work durable. 279 commits.

DEC 1, 2025 — JAN 31, 2026
vendor-portal / checkout
✓
Navigate to checkout00:03 · completed
✓
Fill account details00:08 · completed
!
Obstacle classifiedMFA · human intervention
ThinkReadable, not magical
“A human has more context here.”
I should ask, not guess.
03 / EXTENSION + LOCAL

The agent enters your Chrome.

A local bridge and extension changed the boundary. The agent could work where you were already signed in. The steepest climb: 520 commits.

FEB 1 — MAR 31, 2026
chrome://your-real-session
LOCAL BRIDGE

extension connected
lease acquired
profile your-chrome

No second login.
No synthetic account.
The browser you already trust.
04 / SHARING + MCP

A session becomes a language.

A link could carry replay, screenshots, actions, and context. MCP and CLI parity let different agents read the same story. 280 commits.

APR 1 — MAY 15, 2026
thinkrun.ai/s/a-session-that-speaks
Session.mdCopy for agent
## Reproduction
At 00:17 the pricing card shifts.

## Evidence
• video + cursor
• console + network
• time-synced actions

+ enough context to act
05 / RECORDING PIPELINE

Show, don’t prompt.

Screen, voice, clicks, console, and network became one synchronized artifact. “Watch this” became a complete handoff. 306 commits.

MAY 16 — JUN 17, 2026
Recording · 01:24
Agent watching5 streams synchronized
00:17  “This button jumps.”
00:18  layout shift detected
00:19  console warning attached
00:20  request trace linked

✓ ambiguity collapsed
06 / DISTRIBUTION

The tool learns to travel.

SEO, connectors, remote control, and one pricing source moved ThinkRun toward a product agents can enter anywhere. 148 commits.

JUN 18 — JUL 9, 2026
thinkrun.ai / everywhere
CONNECTORS
● Claude Code
● Cursor
● Cline
● Windsurf
● any agent that reads Markdown

record → understand → act

No mythology.
Just receipts.

Verified from raw Git history and the memory wiki in the July 9, 2026 timeline snapshot.

0
commits
0
merged PRs
0
days of development
0
PRD thread pages
0
commits in peak week · Mar 16—22

What Git cannot remember

The product kept a diary.

Commits say what changed. Memory says why: auth outages, CORS drift, billing pivots, a broken install command, and the patterns that prevented the same mistake twice. The July 9 snapshot drew from 264 memory pages.

The most ThinkRun part of the story

Even the timeline failed review.

The first artifact estimated six numbers, called them verified, then invented a plausible caveat. Independent recomputation caught it.

A product built to make agents observable had to observe itself.

Independent verification · July 9, 2026
Total commits1,5701,571
PRD threads5962
Phase 1 PRs~254
Phase 3 PRs~350147
Strict phase sum—1,571 ✓
VERIFY THE ARTIFACT

This is not the end

The browser finally remembers
what you meant.

Record the problem once. Give your agent the screen, voice, clicks, console, and network—time-synced, inspectable, ready to act on.

See what ThinkRun sees