Find what must not change
Before a single word is removed, TekMyra scans the prompt for spans that carry exact meaning (account numbers, identifiers, dates, legal clauses, quoted terms) and marks them as untouchable for the rest of the run.
Send fewer tokens to your model. Break nothing.
TekMyra is a gateway that sits in the request path between your application and its language model. The open core is the compression library, the benchmark harness and their receipts; the hosted gateway service and the console are the commercial platform, built on the same engine. It shortens the prompt traffic passing through it to reduce what you spend on tokens. When it cannot do that without risking the meaning, it leaves the text alone.
The idea
Prevention over recovery
Most compression tools shorten first and check afterwards. If an account number or a clause went missing, they try to patch it back. TekMyra works the other way around: it decides what must survive before it removes anything, so there is nothing to repair afterwards.
Decline rather than degrade
If TekMyra cannot compress a piece of content and still guarantee those spans survive intact, it abstains. The content passes through untouched, exactly as your application sent it, because a missed saving costs less than a broken span. The benchmark counts a stricter outcome too: an item whose checks fail even after fallback is refused, and every refusal lands in the published numbers at zero saving.
Every run leaves a record
Each run emits a record of what went in, what came out, and what was protected or abstained on. Savings are read out of those records. No figure on this page or in the console is an estimate.
Why this page carries exactly one number A figure is only meaningful if you can trace it back to the records it came from. This page publishes a single measured figure (in the section below) and it travels with its denominators, its refusal count, and a plain statement of what you cannot yet re-run and why. Every other number has to come from run records you generate in your own console.
How it works
What happens to every request
TekMyra is a gateway: your application talks to it the way it already talks to a model provider, and TekMyra talks onward to the model. The hosted gateway service is part of LaconIQ’s commercial platform; the open-source release contains the compression library and benchmark harness that power it. In the field’s vocabulary this is prompt compression (also called context compression), applied in the request path: a chat product and an agentic pipeline go through the same gateway, because the gateway sees only the traffic.
-
It sits in the request path
Point your application at the gateway instead of directly at the provider. TekMyra receives the request, and the model receives whatever TekMyra passes on. Your application code does not change.
-
It protects the critical spans
The incoming prompt is scanned for content whose exact wording matters. Those spans are pinned before any rewriting is considered, so later stages are not allowed to touch them.
-
It compresses only what is safe
The remaining text (the padding, the restatement, the parts that carry no exact obligation) is shortened. Anything the gateway cannot vouch for is left as it is, and the run is marked as an abstention rather than quietly trimmed.
-
It emits a verifiable run record
The gateway writes down what it did: tokens in, tokens out, which route the request belonged to, which spans were protected, and whether it abstained. That record is the source of every number published about TekMyra.
The console
See what you actually saved
The TekMyra Console is part of the commercial platform: a dashboard you open as a local HTML file. It reads the run records the gateway emitted and shows you the arithmetic: tokens in, tokens out, tokens saved. It reads no other source, and it sends nothing over the network. The open repository documents the run-record format; the emitter and this renderer ship with the platform.
Savings over time
Lifetime, this week, and today, so the question “is this worth running?” has an answer drawn from the records.
Attribution per route
Traffic is grouped by the route it came through, so you can see which parts of your application are compressing well and which ones the gateway keeps abstaining on.
The per-item stream
Individual runs, in order, each with its own in, out, and saved figures. Any total on the page can be walked back to the items underneath it.
Local and self-contained
One HTML file, opened from your own disk. No server to stand up, no account, and no upload of your traffic to anyone, including us.
Measured
62.1458% effective byte reduction
Our public-corpus baseline: 62.1458% effective byte reduction over 138 fixtures (2,708,761 → 1,025,379 bytes), with 6 refusals (4.35% of fixtures, 3.55% of bytes). Refusals stay in the denominator and contribute zero saving. The number is not computed only over the cases where compression succeeded.
Bytes, not tokens
This is a byte-basis figure and is not interchangeable with token-reduction ratios. It is our own measurement on our own corpus, not a head-to-head against other compressors, and we will not call it one until we have published one.
The evidence ships
The repository carries the evidence artifact: the figure, its denominators, the corpus file counts, and a sha256 digest for each of the six corpora the baseline tracks (the four public corpora behind this number, plus the synthetic and long-context benchmark corpora). A digest conveys no content. It is a pre-commitment. Five of the six verify directly against this repository; the sixth, the long-context corpus, was recorded over the private tree, and the README states why, and names the test that binds the shipped bytes instead.
What a clone can and cannot re-run
The four corpora behind this number ship in the repository, and a clean clone can verify their fixture bytes and attribution against the committed baseline. Reproducing the rate itself also needs the reference artifact bundle; a clone without it correctly measures a no-op compressor rather than a flattering number.
Trust
No silent learning
The compression model does not train on your traffic. It changes when a new version is released and at no other time, so behaviour stays fixed until you upgrade.
- Versioned releases only A given release behaves the same way today as it did the day you pinned it. There is no background adaptation and no feedback loop from what passes through your gateway.
- Your traffic stays yours Prompts are not collected as training data. The run records the gateway writes are yours, on your disk, for your console to read.
- Open source, so you can check The engine is released under Apache-2.0. You can read the protection rules and the abstention logic, and disagree with them in public if you want to.
- A reference bundle you can re-run TekMyra publishes a reference artifact bundle as a release asset, with its checksum in the release notes. Its results are receipt-verifiable: run it yourself and check the records against what the paper reports.
Who builds this
TekMyra is a LaconIQ product
TekMyra is built and maintained by LaconIQ, and it is the same engine that powers LaconIQ's commercial platform, released in the open. The span-protection approach you read about above began as LaconIQ's answer to a harder question: how do you let an enterprise compress traffic that carries personally identifiable information, contract clauses, and account data without ever losing any of them?
If that is your situation (PII at scale, governance requirements, tenant-scoped deployment, or someone accountable to call), the enterprise side of this engine lives at laconiq.ai. The open-source gateway here is complete and genuinely free under Apache-2.0; the enterprise platform is where LaconIQ adds what companies need around it.
Launch
Open source on August 31, 2026
The repository, the packaged release and the technical paper go live together. The
Python package is tekmyra-core on PyPI; the README leads with the
from-source install, and so do the docs:
install: from a clone, pinned for reproduction
pip install -c constraints-reproduce.txt -e .
install: from PyPI
pip install tekmyra-coreEverything the published numbers rest on ships alongside: the corpora, the baseline evidence artifact, and the scripts that let you check the claims yourself.