Screenshot evidence
00:00Landed on /00:28Opened dashboard01:42Peer-help post02:10Evidence packView Beta
Open the real public dashboard.
Inspect public demonstration runs for GitHub, Self Degree, and this landing page in the separate deployed dashboard. These demos show the artifact format; they are not customer-grade research engagements.
Real dashboard
Three public demonstration runs ready to inspect
- GitHubExample screenshots, peer-help threads, path-crossing events, and raw action timelines.
- Self DegreeExample cohort notes, trust notes, screenshots, and loading flags.
- BetaBotsExample dashboard-visible artifacts, peer help, and path traces.
Evidence timeline
Every run leaves a clocked trail.
See what the bot saw, what it tried, what it waited for, and what changed its mind.
T+00:01Screenshothero screenT+00:11Clickprimary CTAT+00:23Wait2.8sT+00:41Thoughttrust concernT+00:49Peer helpthread startedT+01:06Path tracenear missT+01:31Wait3.1sT+01:55Friction noteimage lazy-load
Quick start
Install the Betabots skill for your harness.
Choose your agent runtime, install the skill/plugin, then ask it to launch a cohort against any product URL.
Install for Codex
git clone https://github.com/yevgeniusr/betabots.git && cd betabots && scripts/install-local.sh codexInstalls the Codex plugin and mirrors the Betabots skill into ~/.codex/skills.
General by design.
Betabots do not ship with product-specific routes, fixtures, or secret backend knowledge. They behave like people in a browser: reading the screen, following visible affordances, getting stuck, asking for help, and leaving evidence.
Screen-drivenNo hardcoded routesWorks across products
What a run leaves behind
Every run writes a complete, inspectable trail.
Open it in any editor. No databases. No lock-in. No hidden black box.
FileWhat is inside
summary.jsonscores, actions, flags
analysis.mdplain-English findings
screenshots/captured visible screens
raw/*.mdfirst-person logs
peer-help.jsonthreads, replies, invites
path-traces.jsonnudges, near misses
timeline.jsonclocked action trail
loading-flags.jsonwaits and unresolved loading states
All yours. Plain files. AGPLv3.
Action timeline
00:00Landed on /00:12Scrolled00:28Opened demo01:29Friction noted
Raw story excerpt
I like the clarity of the value prop and social proof. The pricing page is close, but I need more detail on what is handled.
Path-crossing event
Nudge #12 introduced a pricing concern into a peer-help thread.
Simulated people
Truthful by design.
Truthfulness is default. Not a toggle.
Maya 1
seed-stage founder- Goal
- Protect runway and reputation.
- Constraints
- Time, risk, small team.
- Evidence
- Action trail, visible screens, peer threads
- Truth note
- Sees potential but needs trust signals.
Alex 3
marketer- Goal
- Drive signups without wasting money.
- Constraints
- Clarity, tooling, team buy-in.
- Evidence
- Action trail, visible screens, peer threads
- Truth note
- Confused by pricing and next steps.
Priya 2
student- Goal
- Find a plan she can afford.
- Constraints
- Time, budget, confidence.
- Evidence
- Action trail, visible screens, peer threads
- Truth note
- Wants help choosing the right plan.
Jonah 4
designer- Goal
- Understand if the product respects craft.
- Constraints
- Signal, noise, attention.
- Evidence
- Action trail, visible screens, peer threads
- Truth note
- Needs concrete proof of value.
Peer help
When a bot gets stuck, it asks the room.
Help requests surface fast. Peers reply. Path-crossing nudges can step in. The loop is broken, or the product issue is recorded.
Stuck loopPeer-help postPeer replyPath rescueRecorded
Peer-help thread T-142Date picker confusion
I keep opening the date picker, but selecting a date does not close it. I am stuck in a loop.
Try clicking outside the panel. It closed for me, but it was not obvious.
Confirmed. The loop is recorded and the next action shifts to a clearer surface.
Path crossing
Cross paths, near miss, rescue loops.
Path-crossing traces shape the run so people collide, almost connect, and recover when stuck. Not randomness. Recorded intervention.
Orchestration console
LiveEventBotsTriggerState
cross_pathsMaya 1 + Alex 3align timingdone
near_missPriya 2 + Jonah 4stagger exitsdone
rescued_loopMaya 1open new pathdone
hunchAlex 3surface trust signaldone
Not QA automation.
Betabots does not replace deterministic tests.
Not an analytics SDK.
It captures intent, not just events.
Not scripted praise.
Bots report friction and doubt.
Local-first. Open source. Inspectable.
You are in control.
Contributor proof
Start with the quick start command, inspect the AGPLv3 license, then open the repository when you are ready to contribute.
Runs locally
On your machine and your target URL.
Open source
AGPLv3. Fork it, inspect it, extend it.
Local dashboard
Read artifacts from .betabots/runs.
Real browser sessions
Human-paced browsing, not SDK events.
Raw artifacts
Everything recorded. Nothing hidden.
GitHub
Contribute, report, learn together.
Managed BetaBots review
Begin with a bounded five-BetaBot sample.
One agreed scenario, one cost-efficient model, and concise evidence—subject to eligibility and capacity.
Review the managed serviceSample first. Decide after.
- Free sample5 synthetic BetaBots, one agreed scenario, concise sample findings and evidence.
- Paid deep researchUp to 100 synthetic BetaBots, research-grounded diverse cohorts, a model mix chosen for the engagement, and a detailed prioritized report.
- Commercial stepA custom proposal and quote only after the sample—no hardcoded package price or promised outcome.