Stateful perception for browser agents.
Cut the noise.
Keep the signal.
As humans, we naturally focus on what matters in the moment. Our brains discard what we already know, so our attention lands on what's new.
AI agents have no such filter. Every step, they read the page again: signal and noise, from scratch. JD Codec gives agents that same selective focus. Strip away the known. Surface what matters.
By joining, you agree to receive JD Codec waitlist and early-access emails and acknowledge our Privacy Policy. Unsubscribe anytime.
4 models · 10 tasks · 3 reps · 240 attempts · see the public benchmark →
Same task. Same model. Same tools.
One finishes before
the other starts reading.
Per-snapshot compression from early codec spike tests. Six standalone tasks. Different test from the four-model, ten-task public benchmark below.
Task: Filter products by status 4/6
Same model. Same task. Same tools.
99% smaller page per step.
Benchmarked · 2026-05-14
Same agent. Same model.
Cheaper on every one.
Four models, ten tasks, three runs each. Codec on vs codec off, matched pair. The full task-by-task breakdown lives on the public benchmark.
| Model | Tasks finished | $ per success | Cheaper per success | Smaller LLM bill |
|---|---|---|---|---|
| Gemini 3.1 Pro preview | 12 / 27 (+2) | $0.60 (was $1.27) | 2.13× | -44% |
| Gemini 2.5 Pro | 14 / 27 (+1) | $0.58 (was $0.77) | 1.33× | -19% |
| Gemini 2.5 Flash | 13 / 27 (tie) | $0.052 (was $0.083) | 1.60× | -38% |
| Claude Sonnet 4.6 | 19 / 27 (+2) | $1.06 (was $1.60) | 1.51× | -27% |
| Across 4 models · 216 cost-aggregate attempts | +5 PASS, no losses | 1.3–2.1× | -32% mean |
240 attempts across 10 tasks. 216 feed the cost aggregate above. The tenth task is a privacy probe where the codec correctly redacts a customer email before it reaches the model, reported separately on the public benchmark.
Methodology: WebArena Magento admin lookups, deterministic scoring, provider-native token counts, public list pricing as of 2026-05-14. See the full batch →
Setup
Install. Connect.
Start saving.
Install
Pick your flavour. One command.
npm install -g jdcodec
pip install jdcodec
Requires Node.js 22+. Python-first developers can use pip.
Connect
Drop-in layer between your agent and the browser. No changes to your agent, model, or tools.
jdcodec start
Save
Smaller LLM bill on every model we tested. 1.3 to 2.1× cheaper per successful task. Your agent sees only what changed, and performs better because of it.
Your agents have been
operating blind.
Alpha access is limited to high-velocity teams. Join the waitlist to run your agents at a fraction of the cost, with personal data redacted before it ever reaches the model.
By joining, you agree to receive JD Codec waitlist and early-access emails and acknowledge our Privacy Policy. Unsubscribe anytime.