hunch
Python browser agent where Jev picks each operation and element in about 150 ms, escalating to an LLM only when unsure.
Overview
hunch is a zero-dependency Python browser agent built on agent-browser. Its premise is that most browser steps, such as typing an email or clicking Sign in, are reflexes rather than reasoning. Each step takes an accessibility snapshot (a numbered table of interactive elements), asks Jev in one request which operation to perform and which element to target, runs the action, and verifies the outcome in code.
When confidence is low, the page stops changing, or a click looks irreversible, hunch stops and returns a precise report for a human or an LLM to handle. The README's own benchmark on four page states reports a median of 153 ms per Jev decision versus 678 ms for gpt-4o-mini, both 24/24 correct, and a full form-fill run of 3.4 s with 675 ms spent deciding at about $0.0002. It cautions that this is one simple task and that snapshots and actions add fixed time for any model. Secret values are referenced by name and never sent to TypeSafe.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which operation should be performed next? | TYPE_TEXT · CLICK · SCROLL · ... |
| Choice | Which element should the operation target? | — |
| Yes / no | Is this action risky or irreversible?(inferred) | yes · no |
How it works
- Snapshot the page into a numbered element table
- One Jev request returns operation and target heads
- If confident, perform click/fill/scroll via agent-browser
- Verify progress in code; repeat or exit on success
- If unsure, risky or stuck, exit with a report for a human or LLM
Strengths
- README measures 153 ms median per decision vs 678 ms for gpt-4o-mini
- Decision cost of about $0.0002 for a whole form run
- Success is verified in code, not by the model
- Passwords are passed by name only and never sent to the API
Caveats
- Benchmarks cover one simple task and one network location
- Hard multi-step reasoning still needs an LLM fallback
- Depends on agent-browser and a TypeSafe key
Getting started
`npm install -g agent-browser && agent-browser install`, then `pipx install git+https://github.com/HAR5HA-7663/hunch` and set TYPESAFE_API_KEY.At a glance
- Type
- Project
- Status
- prototype
- Language
- Python
- Runtime
- CLI on top of agent-browser
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
⚡ Browser agent that acts on a hunch: Jev (TypeSafe System One) picks every click in ~150 ms, an LLM is only needed when confidence drops. Zero-dependency Python on top of agent-browser.
hellogumbo/awesome-jev
Frequently asked questions
What does hunch decide?
For each browser step, which operation to perform and which element to act on, plus whether to stop and hand off.
When does it use an LLM?
Only when Jev's confidence is low, the page is stuck, or an action looks irreversible; hunch then exits with a report.
Is my password sent to TypeSafe?
No. Only the value's name is sent, and the README says a test checks this.
Related projects
Chrome extension port of Jev Ultrafast that drives your current tab, with Jev picking the operation and element each step.
Browser extensionbetaTypeScriptBounded browser execution for parent agents: Jev picks the next typed action, agent-browser runs it, stuck states hand back.
ProjectbetaTypeScriptPython decision layer for browser and desktop automation: code observes, acts and verifies; the model makes one typed choice per step.
ProjectbetaPythonManifest V3 Chrome port of jev-ultrafast that drives your current tab, with Jev choosing each operation and element.
Browser extensionbetaTypeScriptReference framework where Jev picks a concrete computer-use action and its execution route from legal, verifiable candidates.
AppresearchPythonAutoGPT agent platform, listed as adding Jev decision blocks for typed routing, filtering, scoring and confidence-gated next actions.
ProjectproductionPython