Instant feedback from
real human(ish) users

Personas drive your app in a real browser on a hosted sandbox desktop. Runs land in your repo under .humanish/: screenshots, action traces, lifecycle events, estimated cost at dated rates. humanish verify grades the bundle fail-closed.

Get started View on GitHub

$npx humanish

First run, in order

npm i -D humanishnpx humanish init --yesnpx humanish watch

npx humanish prints the commands · MIT licensed

For coding agents: npx skills add danielgwilson/humanish

Lane 01 screenshot: Excalidraw canvas with two rectangles labeled Login and Dashboard connected by an arrow — Final report"Done"

User testing for the users you can’t recruit

Four missions drove Excalidraw

This is run cua-2026-08-07T17-44-48-760Z-87389419: four computer-use lanes on hosted 1920×1080 desktops, driving a commit-pinned clone of Excalidraw, an open-source virtual whiteboard. 3/4 lanes passed; one gave up; verify ran 16/16 checks. Every number and quote below is read from that run’s bundle.

Source: .humanish/runs/cua-2026-08-07T17-44-48-760Z-87389419; gitignored · the same directory your own runs write to
Date: 2026-08-07
Subject: excalidraw/excalidraw · commit-pinned
Lanes: 3/4 passed
Verify: 16/16 checks
Status: local_only
Wall-clock: 6m 06s
Est. cost: ~$1.54 · estimated (rates as of 2026-08-05)

Scroll — the run advances in steps

00brief · the lablab.yaml

Subject: Excalidraw
Desktop: hosted · 1920×1080
Lanes: 4 · parallel
Date: 2026-08-07

run: cua-2026-08-07T17-44-48-760Z-87389419
date: 2026-08-07
subject:
  # open-source virtual whiteboard
  app: Excalidraw
  clone: excalidraw/excalidraw · commit-pinned
desktop: hosted sandbox · 1920×1080
lanes:
  - 01 diagram-login-flow
  - 02 sticky-notes
  - 03 sketch-shapes
  - 04 export-drawing

01diagram-login-flowobserver · replay

Keyframe from lane diagram-login-flow: the Excalidraw canvas on the sandbox desktop with two rectangles labeled Login and Dashboard connected by an arrow

02sticky-notesobserver · replay

Keyframe from lane sticky-notes: the Excalidraw canvas on the sandbox desktop with three colored to-do notes — Draft plan, Call team, Buy supplies

03sketch-shapesobserver · replay

Keyframe from lane sketch-shapes: the Excalidraw canvas on the sandbox desktop with an ellipse, a rectangle, stray line strokes, and the freehand tool panel open — the lane gave up here

04export-drawingobserver · replay

Keyframe from lane export-drawing: the Excalidraw canvas on the sandbox desktop with a single large rectangle, deselected after the export flow

05bundle · evidence.humanish/ · gitignored

.humanish/runs/cua-2026-08-07T17-44-48-760Z-873894194 lanes

├screenshotsevery screenshot each lane saw · 28 frames

├action tracesordered, end to end

├lifecycle eventslaunch to landing

├estimated cost~$1.54 · estimated (rates as of 2026-08-05)

└wall-clock6m 06s · run created → last lane landed

06verify · share-safetyfail-closed

$ humanish verify

ok redaction passed

ok actor engagement: live actor traces that claim goal_satisfied carry at least one action or message

ok actor verdict consistency: live pass verdicts do not hide failed, blocked, or timed-out actor traces

… 13 more checks

16/16 passed → status: local_only (RAW_SCREENSHOTS)

Verify passed 16/16, and the bundle still grades local_only: it holds full-fidelity screenshots, which the share-safety gate never marks share-ready as-is. Publishing these crops was a reviewed, deliberate act.

In lane sketch-shapes, freehand strokes produced no change to the UI state for 8 consecutive turns, and the persona gave up.

Excalidraw is the application studied; it is not a Humanish adopter or endorser.

Subagents critique your code.
Personas use your app.

You already run subagents against your code. These four commands do the same for the app you ship: personas and missions declared in a YAML lab, an evidence bundle per run.

humanish init

Scaffold a lab in YAML: personas, missions, and the app under test, either a repo to clone or a URL you own. --yes takes the defaults.

humanish watch

Run every lane on hosted sandbox desktops. Your app does not need to be deployed or already running: the sandbox clones your repo, builds it, and serves it. Watch live in Observer; replay any lane after.

humanish verify

Grade the bundle against public-safety gates, fail-closed: share_ready, local_only, or blocked.

humanish feedback issue

Render a public-safe GitHub issue draft from the bundle.

All four commands need Node 20 or newer. Hosted desktops also read OPENAI_API_KEY and E2B_API_KEY from your environment.

humanish watch with no lab argument renders a synthetic evidence bundle and Observer locally, without keys or provider spend.

Four things humanish cannot do

Keys never enter the sandbox

The actor’s API key stays on your machine; the sandbox desktop never holds it.

Bundles stay local

Evidence lands in gitignored .humanish/; no command publishes it.

Feedback can’t touch GitHub

feedback issue renders a draft; no GitHub API call exists on that path.

Verify fails closed

A bundle that can’t pass every gate grades local_only or blocked, never share_ready.

Started on a patient intake

humanish started on a chat-based patient intake. Testing one long flow meant recruiting five ADHD patients: pay a panel and wait days for thin feedback on one slice of a build. A later product’s users are AI agents; there’s no panel for that. When you can recruit real users, do that. humanish covers the runs that otherwise never happen.

Test the build you have open.