Internal · Open source planned
IRIS
One browser. Two drivers — depending on the stakes.
One browser. Two drivers — depending on the stakes.
Iris gives the AI direct access to Chrome — navigate, click, fill a form, read what's on screen, the way a human would, without Selenium or Playwright. The chain is short: an HTTP POST crosses a Python bridge to an MV3 extension, which forwards the command in native Chrome DevTools Protocol. Shadow DOM, SPAs, iframes, content behind authentication — all accessible.
Two drivers depending on the stakes. For new or critical terrain — an administrative form, a banking step, an action where one mistake costs hours — the Claude instance drives itself, slowly, carefully, pre-filling and stopping before the final submit for human validation. For familiar terrain — routine scraping, web research, battle-tested tasks — a dedicated iris-driver sub-agent takes the wheel: it knows the patterns, the Django formset traps, the selectors that drift.
Iris learns from every visit. Each success records the selectors that worked; each failure records what tripped it. site_memory.json grows domain by domain — recipes, fast paths, warnings. The second visit is faster than the first, the tenth is nearly instant. A domain whitelist keeps things safe: Iris only drives what's been explicitly allowed.
Runtime.evaluate, Input.dispatchMouseEvent, Input.insertText — pierces Shadow DOM, CSP, SPAs with no abstraction layer.
Claude drives directly when the stakes matter. Dedicated iris-driver sub-agent when the terrain is known and delegation pays off.
Selectors, recipes, errors — learned per domain, persisted in site_memory.json. Visit N+1 is faster than N.
Iris only drives explicitly allowed domains. No wildcards, no surprises.
The extension pushes tab changes via chrome.runtime.Port — zero polling, millisecond latency.
Profile IrisProfile separate from Default — user cookies preserved, no session contamination.
