Hindsight

For people who build with Claude Code and Codex

Use hindsight to get better at using AI.

Hindsight reads your Claude Code and Codex sessions next to your git history and works out what you'd do differently: which prompts led to work that lasted, which started dead ends, where the tokens went and which model was worth its cost. It runs on your machine, on the subscription you already pay for.

curl -fsSL https://hindsight.technology/install.sh | bash

macOS and Linux · Free · No account · Runs entirely on your machine · Updates itself · Releases

Hindsight's project view for 3dintelligence: a summary, the lessons with hindsight, and an idea maze of six directions with one dead end and six milestones that held up.

A real project from the maker’s own transcripts: 11 sessions and 232 prompts over two weeks, sorted into 6 directions, 1 dead end and 6 milestones that held up (commits that survived, tests that passed, things you said worked).

Learn from your own history

Advice about prompting and models is written for everyone. Your own transcripts show what worked for you.

Prompt better

Each direction shows the prompt that opened it and how it turned out, whether a milestone in one session or a week-long detour and the sentence of context that would have avoided it.

Spend tokens where they count

Every direction carries its cost in tokens, dollars and hours. See how much went into dead ends, how much of your input came from cache, and which sessions burned the most for the least that survived.

Know which model is better for your work

Benchmarks measure someone else’s tasks. Hindsight measures yours: which model and effort setting reached milestones, and at what cost.

In development

Historical A/B tests: rerun your past with a different choice

Hindsight knows the starting point, the prompt you sent and the result that held up. That makes every milestone in your history a ready-made experiment.

  1. Go back. Check out a throwaway copy of the repository as it stood before the session began.
  2. Change one thing. Replay your real prompt with a different model, a cheaper effort setting, or the sharper prompt Hindsight suggested.
  3. Judge against what happened. Compare each attempt with the change that survived in git, on quality, cost and time.

Experiments run through your own CLI.

Every project you’ve built with agents, on one timeline

Hindsight scans the transcripts Claude Code and Codex keep on your disk and folds worktrees into their repositories. The first scan takes seconds.

  • Every session, placed where it happened
  • Milestones that held up: commits that survived, tests that passed, things you said worked
  • Dead ends: reverted work, abandoned branches, directions you never came back to
Hindsight's projects page: fifteen projects on a June-to-October timeline with sessions, milestones and dead ends marked.

Knowing what you know now

Ask Hindsight to analyze a project and it reads the sessions and the commits together, then writes down what it would do differently a second time.

“Round 0 already showed candidate generation was the bottleneck, so go straight from there to the break-and-fix stage, skipping rounds 2 to 4 of pre-ranker and proposal models (about $800 and a day) that only approached what break-and-fix found in a morning.”What Hindsight wrote for 3dintelligence

Every milestone cites the commit, test or message that backs it.

A table of milestones that held up, each with its date, kind of evidence and the commit or message that backs it.

Token maxing, measured

See what your Claude Code and Codex subscriptions bought you: tokens in and out, what the same work would cost at API list prices, how many times over the subscription paid for itself, plus streaks and badges. Share a card drawn in your browser.

$80,340API-equivalent
$1,600paid in subscriptions
50×API cost over subscription cost

The maker’s own numbers, May to October 2026: 53.8B tokens over 67 active days.

Hindsight's token maxing page: $80,340 at API list prices versus $1,600 in subscriptions, monthly chart and badges.
The token maxing page on a phone.

Your transcripts never leave your machine

Your prompts are the most sensitive thing you write. Hindsight has no server and no account, so nobody else holds them.

Local first

Scanning, storage and the app run on your computer. The database is one SQLite file in ~/.hindsight.

Your subscription, not ours

Analysis runs headless through the claude or codex CLI, on the plan you already pay for, so there is no API key, no per-token bill and no new vendor to get approved.

Nothing phoning home

Hindsight’s only network traffic is fetching its own updates. Your CLI talks to Anthropic or OpenAI as it already does.

How it works

  1. 1

    Install

    One command installs a small native launcher, with no Node or Docker. It fetches the engine, verifies it and keeps it current.

  2. 2

    Scan

    Hindsight finds your Claude Code and Codex transcripts and opens the app in your browser while the scan runs.

  3. 3

    Analyze

    Pick a project. Your own CLI labels its directions, dead ends and milestones, and writes down what to do differently next time.

  4. 4

    Improve

    Carry the lessons into the next project. Optional hooks rescan after every session.

Your next project, with the benefit of hindsight.

curl -fsSL https://hindsight.technology/install.sh | bash

Free · No account · Runs on your machine