Skip to content
Join early access

promptaware — a local memory layer for AI power users

Never lose a prompt that worked.

PromptAware saves your AI prompts and sessions into a private, searchable library on your Mac. Find anything you’ve asked. Reuse it in one keystroke. Your prompts stay on your machine.

Join early accessSee how it works

Free tier · Pro from $8/mo · No card required

  • Local-first
  • macOS · Apple Silicon
  • Private beta 2026
  • No telemetry
Illustrative demo UI · demo data, not live metrics
How it works

Capture. Search. Reuse.

Three mechanics, one loop: AI work stops evaporating, becomes findable, and gets reused instead of rewritten.

  1. 01 / capture

    [claude code]imported 18 sessions · exact tokens14:32

    [clipboard]⌃⌥C “Postgres migration plan, v2…”15:07

    [api]logged anthropic · 12 runs15:40

    The work that normally evaporates gets kept.

    Opt-in, surface by surface: read-only import of your Claude Code and Codex session logs, a browser helper for claude.ai and chatgpt.com, a ⌃⌥C clipboard hotkey that works with any tool, and an API mode if you bring your own key. Capture from the Claude Desktop and ChatGPT desktop apps isn’t live yet — see the support matrix below.

  2. 02 / search

    ❯search “auth middleware”

    7 matches · local index0.03s

    no network calls made—

    Everything you’ve asked, indexed on your Mac.

    Full-text search across prompts and sessions, from a ⌘K command palette or the three-pane workspace. The index lives inside the encrypted local store — a query is a local read, not a cloud round-trip.

  3. 03 / reuse

    ↵reuse → pasted into frontmost app

    ⌘Ccopy · ⌘O open source thread

    prompt v3 · reused ×5family: 3 versions

    One keystroke, not a rewrite from memory.

    Recall the prompt that worked, paste it where you’re working, and keep its history — Versions, Compare, Runs and Variants sit on every prompt, so you can see what changed between the try that failed and the one that landed, and what each attempt cost.

Why this exists

Your best prompt is buried in a chat you’ll never scroll back to.

If you work with AI seriously, you produce more prompt-craft in a week than most people do in a year. Almost none of it survives the session.

  • 01The prompt that finally worked is forty messages deep in last Tuesday’s chat.
  • 02Terminal sessions disappear when the window closes.
  • 03Your prompts are scattered across Claude, ChatGPT, Cursor, Claude Code, notes and clipboards.
  • 04Git only stores the prompts you deliberately committed.
  • 05Most useful AI work is improvised — and improvised work never becomes a reusable asset.
Works with

Exactly what it captures today. Nothing implied.

Capture honesty: which surface each prompt comes from, and — the part most tools skip — whether the prompt text is actually kept. Some surfaces are built to observe cost without ever retaining what you wrote.

  • Claude CodeEarly access
    Imports local session history — prompt text plus exact token counts· opt-in, read-only
    Kept
  • Codex CLIEarly access
    Imports local session history — prompt text plus exact token counts· opt-in, read-only
    Kept
  • In-app composerEarly access
    Write in PromptAware — cost preflight before you send
    Kept
  • API mode — Anthropic / OpenAIEarly access
    Send with your own key — provider-reported tokens recorded· opt-in
    Kept
  • Clipboard hotkeyEarly access
    ⌃⌥C captures from any tool, anywhere
    Kept
  • claude.ai / chatgpt.comIn progress
    Browser helper reads the prompt box on those sites only· saves drafts as you type
    Kept
  • Claude Desktop / ChatGPT desktopNot yet supported
    No capture yet — the groundwork is in, nothing is recorded· zero-recall by design
    Never kept

“Early access” = ships in the first private-beta build; invites open in batches during 2026 — nothing here is publicly live yet. “Never kept” means no prompt body is stored from that surface, ever. For the desktop apps that currently means nothing is stored at all: the capture path is built but deliberately retains nothing until it can prove it caught a real submission, so treat it as unsupported for now. Editor extensions (including Cursor), Gemini and Antigravity are not in the first release.

Cost intelligence

Know what a run costs before you send it.

Recall is why you’d open PromptAware. The money view is why you’d keep it open — a predicted range, a ceiling that can’t be exceeded, and a provenance mark on every number so you know which ones are measured and which are modelled.

Illustrative demo UI · demo numbers, real vocabulary

  • cap

    One number that isn’t a guess

    Alongside the estimate sits the model’s hard cap — your prompt tokens plus the most output the model is allowed to produce, priced out. Real cost is always at or under it. When a model’s output limit isn’t published, the app shows no cap rather than inventing one.

  • provenance

    Every figure says where it came from

    Verified means the provider reported it. Reconciled means several sources saw one run and it was counted once. Estimated means modelled locally. Mix them in a total and the total downgrades to the weakest source — no laundering an estimate into a fact.

  • basis

    Spend vs usage value

    On a subscription, API-rate arithmetic isn’t your bill. So money splits in two: spend is what actually bills you, usage value is what the same work would have cost at API rates. Switching the basis changes the labels, never the numbers.

  • pricing

    Prices that update without a release

    Rate tables arrive over a signed channel and are verified against a pinned key before they’re trusted — a bad file can never displace the built-in table. Every row carries the date it was true, and goes visibly stale after 60 days.

The prediction calibrates on runs you send through the app’s own API mode — until it has your history the likely-range stays deliberately wide rather than falsely confident. Imported session logs carry the provider’s own token counts, so those costs are measured, not predicted.

Privacy & local-first

Trust shouldn’t require faith. So here’s the architecture.

PromptAware holds your most sensitive work-in-progress, so the privacy model is structural, not a policy promise.

storage

Local encrypted store

Prompts and sessions live in an encrypted database on your Mac, with the key held by the macOS Keychain rather than a plaintext file. There is no cloud copy of your prompt content — there’s nothing to breach on our side.

telemetry

No telemetry. No account.

The app ships with no analytics SDK, no crash reporter and no install ID. Its complete list of network destinations is printed inside the app and enforced by a test that fails the build if the code ever reaches a host that isn’t on it. (This website’s analytics track clicks and pageviews only — never your email or form answers.)

permissions

Opt-in, scoped permissions

Nothing is captured until you switch a surface on, one surface at a time. The browser helper reads the prompt box on supported sites — not your cookies, not your other tabs — and talks only to the local app.

purge

One-click purge

Erase everything PromptAware holds, whenever you want — and get back a receipt listing exactly what was deleted and what was re-checked as gone. Your library, your call.

waitlist

The waitlist needs one thing

An email address. Every other question on the form is optional and exists only to prioritise beta invites.

boundary

The precise boundary

Prompt content leaves your machine only when you send it to a provider yourself, with your own key. Rewrites run on a small model on your Mac, so they send nothing — the only thing ever downloaded is the model file. Anonymous prediction stats are off unless you turn them on, and carry numbers only.

Pricing

Pay only if it earns its place.

No checkout exists yet — this is intended pricing, and everything is free during the private beta. Telling us your intent on the waitlist is how founder seats get prioritised.

Free

$0

Useful forever for trying PromptAware.

  • In-app composer
  • Cost estimate before you send
  • Live token counts
  • A starter prompt library
  • Clipboard hotkey capture
Join early access

Pro

Most popular
$8/mo

For people who live in these tools daily.

  • Unlimited local prompt history
  • Full search & recall
  • Version families & diffs
  • Reuse metrics & export
  • Spend & usage-value analytics
Join early access

Founder Lifetime

First 100 · pre-launch
$129once

Pay once, keep Pro for life.

  • Everything in Pro
  • Lifetime updates
  • Founder badge & direct input
Join early access

Founder seats

100 founder seats. When they’re gone, lifetime goes away — Pro becomes $8/mo for good. Available only during the pre-launch → launch window, then retired permanently.

Early-bird annual

Join before launch and lock in $60 for your first year, versus $96 standard annual.

Nothing is charged today and joining the list commits you to nothing. Founder-seat and paid-intent signups are first in line for the private beta.

Early access

Stop losing the prompts that work.

Join the waitlist and tell us whether this is just interesting — or something you’d actually pay for. That answer decides the invite order.

How interested are you?
›Help us prioritise your invite (optional)
Which do you use most?

No spam. We’ll only email you about PromptAware early access and product updates.

Founder-seat and paid-intent signups get the first beta batches

FAQ

Direct answers to the real objections.

What is PromptAware?
A local memory layer for AI power users. It’s a macOS app that captures the prompts and sessions that normally disappear, organises them into a private, searchable library on your Mac, and puts the prompt that worked one keystroke away.
Is this a chatbot?
No — there’s no assistant to talk to and no cloud model behind it. The job is remembering: capture, search, recall. It can also rewrite a prompt you’ve saved into four variants (concise, detailed, structured, creative), and that generation runs on a small model on your own Mac. It stays switched off until you choose to download the model, and no prompt text is sent anywhere to do it.
Does it replace Claude, ChatGPT or Cursor?
No. It sits alongside the tools you already use and keeps the work you produce in them. You change nothing about how you prompt.
Which tools does it actually support today?
Read-only import of Claude Code and Codex session logs (prompt text plus the exact token counts the provider reported), the in-app composer, an API mode for Anthropic/OpenAI with your own key, a ⌃⌥C clipboard hotkey that works with any tool, and an opt-in browser helper for claude.ai and chatgpt.com. The Claude Desktop and ChatGPT desktop apps aren’t captured yet — the groundwork is in the build, but it records nothing until it can prove it caught a real submission. Editor extensions, including Cursor, aren’t in the first release. The support matrix above is the source of truth, and we keep it honest.
Will it read all my prompts?
No. Nothing is captured until you turn it on, one source at a time — there’s no monitoring running in the background. The browser helper reads only the prompt box on the sites you’ve enabled: not your cookies, not your other tabs, not the rest of the page. Importing session logs only reads them, never changes them, and any API keys or access tokens it recognises are removed before anything is saved. You can also set it to record what a session used without keeping what you wrote.
Does my prompt content leave my Mac?
No. Your prompts sit in an encrypted database on your Mac. The app has no analytics and no crash reporting of any kind, so there is nothing that could carry your prompt text out. It leaves only when you send it to a provider yourself, with your own key. Rewrites run on your Mac too, so they send nothing. There is one thing you can switch on: anonymous numbers showing how close the cost predictions were. It is off unless you turn it on, it is only numbers, and it cannot include your prompt text or anything that identifies you.
Why not just use git / CLAUDE.md?
Git holds the prompts you deliberately committed. PromptAware keeps the ones you improvised — the terminal session, the late-night chat, the twelve variations before the one that finally worked. Your repo prompts can be imported later; your everyday AI exhaust should not disappear.
I use skills and slash commands, not saved prompts. Is this for me?
Yes. When we say prompt, we mean anything you type to an AI — a skill, a slash command, or a sentence you typed once in a session and never wrote down. Skills are the ones you liked enough to keep. Most of what you type never gets that far: you explain something, it works, and it’s gone when you close the window. PromptAware keeps those too, along with the session they came from. You don’t have to save anything or change how you work — it captures as you go. And when you notice you’ve typed roughly the same thing a few times, that’s usually a good candidate for your next skill.
I never look back at old prompts. Why would I start?
Because today there’s nowhere to look. When you need something you wrote last month, typing it again from memory is faster than hunting through old sessions — so that’s what everyone does. Retyping is the cost nobody notices. If it’s all in one place you can search, finding it takes seconds, and it stops being the slower option. Even if you never search at all, the app still shows what each session cost, and what a run is likely to cost before you send it.
Why not use a free token tracker?
A tracker tells you what you already spent. PromptAware does that too, but it also keeps the work: every prompt and session, searchable, so you can find what worked and use it again. And it shows you the cost before you send rather than after, while you can still change your mind. Finding things again is why you’d open it; the money is why you’d keep it open.
What happens if browser capture breaks?
Sites change their markup; capture helpers break. When that happens the surface is marked degraded in the app — honestly, not silently — and the clipboard hotkey keeps working everywhere. Your existing library is local and is never at risk from a broken capture path.
Is this for teams or individuals?
Individuals first: your prompts, your library, your machine. Team features may come later, shaped by what early users actually need.
When does early access open?
Private beta invites go out in small batches during 2026, starting on macOS (Apple Silicon). Founder-seat and paid-intent signups from the waitlist are first in line.
PromptAware — Never lose a prompt that worked