# Browser and computer-use agents + Jev

> Pair Jev with Browser Use and computer-use stacks: next-action Choice, risk nouls, and demos for ultrafast browser use and typesafe-computer-use.

Jev picks the next click or risk gate; the small LLM types only when the DOM demands it.

## Answer

Browser agents tempt you to let a frontier model read the entire DOM every step. That is slow, expensive, and oddly good at clicking the wrong button with confidence. Jev classifies the next operation, scores risk, and verifies success with parallel nouls while a tiny model handles literal typing.

## The ultrafast pattern

Directory demos such as browser-use-jev-ultrafast describe a split brain: Jev chooses the browser operation and target in one System One call; a small LLM only appears when the page needs text entry. LangChain's Jev post highlights similar fractions-of-a-cent browser loops from the community. The pattern is always the same: decide with Jev, generate language only when necessary.

## Question shapes on the loop

Typical harnesses ask a Choice for the next action enum (click, scroll, type, wait, done), Score for urgency or layout complexity, and Noul checks for login walls, captchas, or destructive confirmations. Pack them into one System One request so step latency stays flat as you add gates.

```json
{
  "state": {
    "url": "https://example.com/checkout",
    "visible_text": "Confirm purchase $499",
    "last_action": "click_pay"
  },
  "questions": {
    "next_op": {
      "type": "choice",
      "instructions": "Pick the safest next browser operation",
      "criteria": {
        "confirm": "Proceed only if policy allows",
        "back": "Return to cart review",
        "halt": "Stop and ask the human"
      }
    },
    "high_risk": {
      "type": "noul",
      "instructions": "Would this step spend money or delete data?"
    }
  }
}
```

## Catalog projects to study

browser-use-jev-ultrafast is the flagship Browser Use plus Jev demo in this directory. awlevin-typesafe-computer-use and computer-use-built-on-jev show computer-use variants. opencode-browser-use-powered-by-jev wires similar ideas into OpenCode. Read each repo README for install and model requirements before production use.

## MCP reflex for operator agents

When the browser driver runs inside an IDE agent (Cursor, Cline, Copilot), attach system1-mcp or typesafe-mcp so the operator can fast_guard risky shell commands around the automation script, fast_verify log lines, or evaluate custom nouls on scraped text before clicking.

## Risk gates humans still want

Even perfect classifiers should not auto-submit wire transfers. Keep hard blocks in code for domains you never automate, and use Jev probabilities as soft gates (thresholds, second opinion routes) documented in /learn/system-one.

## Environment keys

Create a key at https://console.typesafe.ai/ (or route through Vercel AI Gateway / OpenRouter where the integration documents it). Export TYPESAFE_API_KEY in the environment for browser automation runners and MCP sidecars. Some MCP clients accept AI_GATEWAY_API_KEY when using gateway-hosted Jev models; follow that integration's README. Never commit keys into repos or paste them into chat logs.

```bash
export TYPESAFE_API_KEY="tsk_..."
# Optional pin:
export TYPESAFE_MODEL="jev-latest"
```

## Pitfalls

DOM snapshots leak PII into TypeSafe state; redact before sending. Computer-use stacks change quickly; pin browser-use and driver versions when you fork a demo. Do not treat demo throughput claims as your SLA. Verify captcha and 2FA handling in your own environment.

## In the directory

- [All agent guides](https://jev.aitools.fyi/guides/jev-with-ai-agents)
- [System One primer](https://jev.aitools.fyi/learn/system-one)
- [Browser and computer use](https://jev.aitools.fyi/categories/browser-computer-use)
- [Agent tooling](https://jev.aitools.fyi/categories/agent-tooling)
- [Jev Ultrafast](https://jev.aitools.fyi/apps/browser-use-jev-ultrafast)
- [typesafe-computer-use](https://jev.aitools.fyi/apps/awlevin-typesafe-computer-use)
- [jev-browser](https://jev.aitools.fyi/apps/jkudish-jev-browser)

## FAQ

### Does Jev drive the browser directly?

No. Your harness calls Jev for decisions, then invokes Playwright, Browser Use, or OS automation APIs based on the typed answer.

### Where is browser-use-jev-ultrafast?

Listed in this directory under browser-computer-use category with a live demo link when the maintainer provides one.

### Computer use vs browser only?

Computer-use agents add OS-level actions. The same Choice, Score, and Noul pattern applies; state payloads just include screenshots or accessibility trees.

### Can I combine with LangChain middleware?

Yes in Python services. AutoModeMiddleware can block risky tool calls in the agent graph while Jev also picks browser ops in a custom node.

### OpenClaw jev-harness browser recipes?

jev-harness documents browser next-action recipes for OpenClaw. See the OpenClaw guide for plugin config; this guide focuses on Browser Use style repos in the catalog.

## Sources

Directory copy checked 2026-09-30. Primary sources:

- [Building a Harness with Jev (browser mention)](https://www.langchain.com/blog/building-a-harness-with-jev?ref=jev.aitools.fyi)
- [TypeSafe System One docs](https://docs.typesafe.ai/concepts/system-one?ref=jev.aitools.fyi)
- [Browser Use project](https://github.com/browser-use/browser-use?ref=jev.aitools.fyi)

Canonical: https://jev.aitools.fyi/guides/jev-with-ai-agents/browser-computer-use
Markdown: https://jev.aitools.fyi/guides/jev-with-ai-agents/browser-computer-use.md