Trust

Your PC runs the work. We handle the license.

PromptParle is an AI context optimization gateway that sits between your desktop tools and providers such as OpenAI, Claude, Gemini, and Grok. It strips bloated tokens, keeps the useful signal, and routes live prompts so you can use flagship models at a lower effective cost. Desktop client 0.25+ optimizes context and calls your model on your machine.

The data path

Prompt, context, and provider API keys stay on your PC. The desktop client optimizes locally, then calls OpenAI / Claude / Gemini / Grok with a key stored only on that machine (DPAPI on Windows). PromptParle cloud is not on the model path.

1Your PC, secret gate, dial/profile optimize, drop journal, workspace / Git / SSH tools.
2Your provider. HTTPS direct from the desktop with your local BYOK key. They bill the tokens.
3PromptParle portal , account, plan, desktop license key (pp_live_), client install package. No prompt bodies. No provider keys.

Two different keys

Desktop key pp_live_…, proves this PC may use your license. Stored locally (DPAPI). Hash only on the server.
Provider keys sk-… / Claude / Gemini / Grok, stored only on the PC. Set with Set-PromptParleProviderKey or Providers in the local UI. Never uploaded to PromptParle.

Secret gate (on the PC)

Credential-shaped patterns (API keys, tokens, PEM blocks) are masked on your machine before any provider call. Default policy is strict: residual high-confidence matches block the send. This is not DLP for IPs or hostnames, under local-first those only go to the model provider you chose, not to PromptParle.

Dial 1-5 & savings expectations

Savings depend on the dial and how noisy the pack is. We are still measuring real workloads; published examples (Noisy ~78%, Security ~60%, Clean ~2%) show the shape of results so far. Clean unique prose often barely moves. Noisy logs, security packs, multi-file dumps, and agent chains are where reduction usually shows up. Proof lives in the dial savings line in the desktop UI per turn.

Dial	Name	What it does	Typical reduction
1	Max Fidelity	Keep nearly everything. Lightest touch, for reviews where every line may matter.	0-10%
2	High Fidelity	Trim obvious junk only: clear duplicates, chrome, and pure filler.	5-20%
3	Optimized	Balanced default. Collapse low-signal bulk while protecting the ask and unique evidence.	15-45%
4	High Savings	Aggressive on noisy logs, multi-file dumps, and repeated frames.	40-70%
5	Max Savings	Maximum collapse of low-signal material. Best for fat packs you already trust the profile on.	50-80%

Bands assume mixed or noisy work. Clean unique prose often sits near zero at any dial, that is correct. Use dial 1-2 when every line may matter. The UI can show a drop journal (what was collapsed). Provider prompt-prefix caching is complementary for stable system/docs, we are not a response cache.

See published example packs.

Heuristic categories (open book)

Deterministic pruning, not an LLM summarizer. Categories below; full scoring stays product craft.

Repetition & near-duplicates

Identical or near-identical lines, blocks, and frames get collapsed with a count, especially logs and copy-pasted snippets.

e.g. Heartbeat / health-check spam · Repeated stack frames · Same paragraph pasted twice

Boilerplate & chrome

License banners, tool headers, decorative separators, and install fluff that rarely change the answer.

e.g. LICENSE / NOTICE blocks · ASCII banners · Generic “generated by” headers

Structure over raw bulk

Prefer outlines, signatures, and error neighborhoods over entire files when the ask is local.

e.g. Function signatures vs full implementations · Error ± context window vs whole log day · Config keys that matter vs every default

Profile bias (job-shaped keep lists)

Each profile tips what to keep: security-review leans auth and trust edges; log-analysis leans events and anomalies; developer leans code structure.

e.g. security-review → auth, secrets, egress · log-analysis → errors, unique events · documentation → headings and claims

Dial 1-5 (fidelity ↔ savings)

Lower dials keep more original text. Higher dials allow more aggressive collapse of low-signal material. You choose the tradeoff per turn.

e.g. Dial 1 · Max Fidelity · Dial 3 · Optimized (default) · Dial 5 · Max Savings

Secret masking (not compression)

Credential-shaped patterns are masked on your PC before the model call. Safety, not a savings trick. No scanner is perfect.

e.g. API keys · Bearer tokens · Password-like assignments

What we try hard not to “optimize away”

Your actual question, unique prose, rare identifiers, and the evidence needed to answer, especially under lower dials.

e.g. User ask / constraints · Unique IDs and hostnames you named · One-off error messages

Invite a friend

PromptParle is free and open. Invitations are just a friendly way to bring someone along.

Anyone can create a free account

No code, no waitlist, no gatekeeping. Sign up with email and password (or Google / GitHub), and you're in. Each desktop still gets its own pp_live_ license key.

Invitations are a nicety, not a requirement

Want to bring a teammate or a friend? Send them an invite from your account. It's a warm hand-off, not the only door in, they can also just sign up directly.

Support the project if it helps you

PromptParle is free. If it saves you real money on tokens, an optional pay-what-you-can donation keeps the gateway boringly reliable and the roadmap moving.

The doors are open: create a free account, make a desktop license key, install, and see your savings. Invite a friend when you're ready, no code required.

Create free account

Pricing · FAQ · Install