The thesis

Saving tokens is nobody's job. So we made it ours.

Cheaper tokens sit in a blind spot. The frontier labs optimize capability, and more usage is their business, not less of it. Developers chase the next model and the next feature, not last month's bill. So the savings problem sits unowned, in no man's land, and the partial fixes that do exist do not add up.

The repo treadmill

Every week a new viral repo: a cache trick, a router, a compressor. You bolt it on and move on when the next one ships. Each is brilliant alone, but a pile of bolt-ons is not a scalable model, and they never knew about each other.

Invoices don't change behavior

A monthly total never changed anyone's habits. You change how you work when you see the saving happen, in the moment, inline, as you type. A number in a report you never open is not a feedback loop.

The cold-cache drain

Cache goes cold. Every restart pays to rebuild the same context, the same files, the same prompt. You are not doing more work. You are paying for the same warmup again, and nothing tells you it is happening.

BeforePoint solutions, each paddling its own way

→

A synchronized rowing crew pulling in unison

AfterOne crew, every stroke in sync

So the answer was never one more optimizer. It was a foundation: measure everything, apply the wins automatically, and coach the habit while the work is happening. Prefex is the first product built on that base.

The foundation

Instrumentation. Auto-apply. Coaching.

Three layers sit under everything PromptForce builds. Saving tokens is the visible result. This is the base that makes it compound, and the reason a proxy turns into a platform.

01 · Instrumentation

Every request measured and attributed per mechanism, in honest, incremental terms. You cannot improve, or trust, what you cannot see. Instrumentation came first; the savings followed the data.

02 · Auto-apply

The wins apply themselves. No workflow change, no config archaeology, fail-open by default. A saving that takes effort to claim does not get claimed, so we claim it for you.

03 · Coaching

Savings shown inline, in the moment, so behavior actually changes (you saw it in action up top). The layer that compounds: a number you watch happen sticks.

Prefex runs on this base today. ReadyBase and the tools after it build on the same three layers, which is why this is a company, not a feature.

How we build it

Three non-negotiables.

1 · Fail open

Internal errors never block a request. On bad input, a missing dependency, or a slow path, we log and forward upstream unchanged. The optimization is never a single point of failure.

2 · Binary, not source

One CGO-free Go binary. No runtime fan-out of processes, no sidecar to supervise. Trained weights are the one exception, and they stay behind an opt-in, default-off seam.

3 · Your keys stay home

Everything runs on localhost. No telemetry, no analytics, no external logging. The dashboard has no auth because nothing can reach it but you.

Lineage: ideas we learned from, rebuilt natively.

We credit the sources we distilled from even when not legally required. Copyright protects code, not ideas, so we study the source, extract the intuition, and re-author it inside our own principles.

RouteLLM → explainable rules router headroom / Kompress → cache-safe compaction LLMLingua-2 → token-importance trimming caveman → filler-stripping for text Mem0 → cross-session typed memory MemGPT / Letta → tiered context management OpenHands → task-span focus notes RTK → tool-output compression ModernBERT → on-device routing encoder NVIDIA Switchyard → cascade routing signals (calibrating) Storybloq → structured session handover codeburn → setup-health waste detection Invariant Scanner → tool-call policy taxonomy DefenseClaw → result + admission scanning OPA → embedded policy engine (linked, Apache-2.0) Anthropic native → prompt cache + context editing