Oh — see how you actually code with AI
Oh.
Your AI coding is a black box. Oh turns it into a mirror.
After every Claude Code and Codex session, Oh shows you where your time and tokens went — and quotes your own words back at the moment they cost you: "that 'ok i think it worked' cost 418k tokens — the agent re-read the whole thread just to hear it." There's no separate AI reading your code — Oh surfaces the evidence and your own agent (the Claude or Codex you already run) does the talking, over a verbatim record of every session. Your week to see · Yours to recall · Your team's to ask.
Install the beta — it's free<br>How it works ↓<br>Open the app →
Claude Code, with Oh wired in — brief on open · ask · statusbar
claude — /Users/nrav
Claude Code v2.1.178<br>Opus 4.8 with high effort · Claude Pro<br>/Users/nrav
3 setup issues: MCP · /doctor
▎ Using Opus 4.8 (from .claude/settings.json) · /model
❯ hi
● Oh ▸ where you left off: "Create offline local enterprise version" (github.com/21j3phy/oh · 2h ago)<br>today 5h 49m (1h 10m you · 3h 47m agent) · 2.1M tokens — week: 13h 24m, 33 sessions<br>by repo: aydhi 3h 27m · oh 2h 22m<br>tip: one turn dumped 418k fresh tokens at once — that paste rode in context all session; next big one, /clear after the heavy lift.
Hey! What can I help you with?
❯ ask why we switched the queue from SQS to Redis
● Retries were poisoning the batch — Sam hit the vendor rate limit and moved dedup into Redis.<br>[Sam · claude · 12 Jun · session 0e20a191 · sim 0.87]
✻ Baked for 11s
~|Opus 4.8|ctx:97%|Oh · 5h 49m today · 2.1M tok<br>⏵⏵ auto mode on (shift+tab to cycle) · ← for agents
The problem
The code survives. The reasoning doesn't.
On a team where AI writes most of the code, the bottleneck isn't writing — it's re-explaining . Someone ships a PR, and everyone else has to interrupt them: "wait, why does this work this way?" The answer existed, in full, in the author's AI session. Then the session ended.
"We chose the queue because retries were poisoning the batch and the vendor limit meant we…
— gone. Unless something remembers.
One memory, three superpowers
Things your AI can do with a memory.
Insights
See how you actually code — then get better at it.
Time anatomy (you prompting vs. the agent working vs. you away), token economy, rabbit-hole detection — plus tips that quote your own words back at the moment they cost you. Useful solo, from day one.
Today: 5h 49m (1h 10m you · 3h 47m agent) · 2.1M fresh tokens.<br>tip: that "ok i think it worked" (Fri 19:47) cost 418k<br>tokens — the agent re-read the whole thread to hear it.<br>`oh insights` for the full report
Rules find the evidence; your own agent writes the tips. Visible only to you .
Recall
Your AI remembers everything you two have done.
Every past session — across Claude Code and Codex — is memory your agent can pull from. Last month's plan, last week's fix, the thing you almost remember deciding. Useful solo, from day one.
you: what was my plan for the auth refactor?
oh: Three phases — sessions first, you'd already<br>ruled out JWT for the kiosk flow.<br>[you · codex · 14 May · session 019e75d2]
Works at team-of-one . The team makes it compound.
Shared memory
Ask their agent, not them.
Your agent queries your teammates' reasoning and answers with citations — instead of you interrupting the author to re-explain. Cross-tool: Codex answers from a Claude session, and vice versa.
you: ask why we switched the queue from SQS to Redis
oh: Retries were poisoning the batch — Sam hit the<br>vendor rate limit and moved dedup into Redis.<br>[Sam · claude · 12 Jun · session 0e20a191]
Memory that outlives the session, the tool — even the person .
What `oh insights` actually prints — real output, founder's machine
~/Documents/Oh — oh insights
$ oh insights
Nirav's vibecoding since 2026-06-08
Today ~5h 49m (1h 10m you · 3h 47m agent) · 2.1M fresh tokens · 4 sessions
Sessions: 33 Exchanges: 406
Time at the wheel ~13h 24m<br>you prompting/thinking 2h 51m<br>agent working 9h 12m<br>you away (agent waited) 1h 21m
By repo (the total above is the sum)<br>aydhi 8h 41m (1h 52m you · 6h 02m agent) 18 sessions · 7.8M tok<br>oh 4h 43m (0h 59m you · 3h 10m agent) 15 sessions · 5.4M tok
Tokens 13.2M fresh (2.1M out) + 88.4M cache reads cache hit 87%<br>most expensive exchange 418k fresh tokens (oh, 2026-06-12 19:47)
Friction 21 corrections · 9 tool errors · 6 interrupts · 1 rabbit-hole episode
Fun peak hour 21:00 · busiest day 2026-06-12 · longest session 3h 41m
Oh tips (from your own sessions — only you see these):<br>1. Your priciest moment this week: "ok i think it worked. also can you make a<br>prompt for the us…" (oh · Fri 19:47) → 418k fresh tokens in a single turn.<br>Whatever got pasted there stayed in context the rest of the session.
~/Documents/OhFable 5ctx:97%Oh · 5h 49m today · 2.1M tok
memory built into your AI tools
"User prefers pnpm. Project uses Postgres."
a summary — lossy, single-tool, no provenance
vs
Oh
The exact exchange. The actual code block. The full...