Every time your AI agent looks at a page, it reads the whole thing. Not just what changed — everything. Again. On every step.

For complex web tasks that means tens of thousands of tokens per action, compounding across a session. Costs climb. Failure rates follow. Your agent is spending most of its budget on information it already has.

JD Codec is a stateful compression layer that sits between your AI agent and the browser. Instead of sending the full page state on every step, it tracks what changed and sends only the delta. Your agent gets a precise, minimal view of the current state, without needing to know JD Codec is there.

No changes to your agent. No changes to your model or tools. Drop it in.

In real-world tasks, agents running JD Codec delivered:  
  → 86% fewer input tokens (1,689k → 230k across 6 real-world tasks)
  → +17 percentage points on task success rate (66.7% → 83.3%)
  → 4.6% faster wall-clock execution

JD Codec is in early access. If you are building AI agents for browser automation and spending more than you should on tokens, we would like to hear from you.

jdcodec.com

JD Codec

Org chart

Teams

Offices

JD Codec

Org chart

Teams

Offices

Related companies