TFBthumb  ·  Playground

Browser agents that don't flake. Website audits that don't lie.

Two modes on one substrate. Drive a page by accessible name and the substrate handles the selectors, the timing, the retries, the identity, and the security. Audit any URL end-to-end and seven lenses surface the real bugs: dead buttons, sub-44 tap targets, broken settles, occluded affordances, aria-stuffing, viewport-only loss of functionality. One CLI, one report, three viewports.

The way you'd write it without TFBthumb

Wikipedia changes the DOM under your feet. Its Vue-based Codex search component swaps the bare <input> for a wrapped variant the instant focus lands; the same name resolves to two different inputs (one visible, one hidden for responsive); the success state is in <title> and <h1>, not an aria-live region. The same shape exists on every Material UI, Radix, Chakra, and modern design-system page. Here is what those four problems cost you side by side.

import asyncio
from playwright.async_api import async_playwright

async def search_wiki(query):
    async with async_playwright() as p:
        b = await p.chromium.launch()
        page = await b.new_page()
        await page.goto("https://en.wikipedia.org/")
        await page.wait_for_load_state("networkidle")
        # Which input? Two have aria-label "Search Wikipedia".
        # The visible one is the first; the hidden duplicate
        # would error on .fill(). Hope nothing re-renders.
        box = page.locator(
            'input[aria-label="Search Wikipedia"]'
        ).first
        await box.click()
        await asyncio.sleep(0.05)  # Codex swap on focus
        # The box you held is now an orphaned node. Re-resolve.
        box = page.locator('.cdx-text-input__input').first
        await box.fill(query)
        await page.get_by_role("button",
            name="Search").click()
        await page.wait_for_url("**/wiki/**")
        # Did it work? Parse the title yourself.
        title = await page.title()
        return query.lower() in title.lower()
        await b.close()

curl -X POST https://projecttfb.com/api/v1/runs \
  -H "Authorization: Bearer tfb_beta_..." \
  -H "Content-Type: application/json" \
  -d '{
    "url": "https://en.wikipedia.org/",
    "fields": {"Search Wikipedia": "Pizza"},
    "submit_target": "Search",
    "success_text": "Pizza"
  }'

# Returns:
# {"run_id": "run_...", "status": "queued", ...}
#
# Poll GET /runs/{run_id}:
# {
#   "status": "done",
#   "final_summary": "observed 'Pizza'",
#   "trace": [
#     {"outcome":"acted", "action":"type",
#      "target":"Search Wikipedia"},
#     {"outcome":"acted", "action":"click",
#      "target":"Search"},
#     {"outcome":"done", "detail":"observed 'Pizza'"}
#   ],
#   "ledger_receipts": [...]
# }

Try it right now

Paste a URL, describe what you want done, click Run. The agent runs in a sandboxed headless Chromium on our server. The playground tier refuses every consequential action (submit, send, delete, pay) at the wrapper, exactly the boundary you'd want for an evaluation. The beta opens the human-in-the-loop gate when you're ready to actually submit.

Audit any URL end-to-end

The Inspector runs seven lenses across three viewports (mobile, tablet, desktop) and emits one structured report. Each finding names the affordance, the lens that surfaced it, the severity (BLOCKER, FAIL, WARN, NOTE), the evidence, and a one-line suggested heal. Real bugs from the canonical run set, none cherry-picked:

One CLI call. One report.

python3 tfbthumb_inspect.py https://your-site.com
#  -> 7 lenses x 3 viewports, perf + canvas pixel metrics, evidence screenshots
#  -> structured findings: severity / lens / affordance / observation / heal
#  -> exit 0 (clean) | 4 (BLOCKER) | 5 (journey step failed) | 6 (gate failed)

# Restrict to a subset:
python3 tfbthumb_inspect.py https://your-site.com --lens tap_target,reach,dead_button

# Mobile only:
python3 tfbthumb_inspect.py https://your-site.com --viewport 390x844

# JSON for CI:
python3 tfbthumb_inspect.py https://your-site.com --json > report.json

The lantern, specifically, lit on a dark scene, on mobile and desktop, gated end-to-end

v0.3.3 drove past the title screen. v0.3.4 healed the doctrine-pivot footnote: the customer can now say the lantern, specifically, not "the canvas has any bright spot somewhere." --canvas-roi declares a named rect inside the canvas; --check canvas ... in NAME reads metrics from that rect specifically. Plus reachable-any for the responsive nav pattern (Main Menu visible as a desktop button, as a hamburger on mobile, both honest). The Lanternlight-shape customer writes one invocation and gets one exit code:

python3 tfbthumb_inspect.py http://localhost:3001/ \
    --viewport 390x844 \
    --click "begin" \
    --wait-for "Wonder Book" reachable \
    --canvas-roi lantern 280,140,60,60 \
    --canvas-roi sky 0,0,400,100 \
    --gate "milestone-0.1" \
    --check "Wonder Book" reachable \
    --check "Main Menu" reachable-any \
    --check perf fps>=55 \
    --check canvas brightness>=0.6 in lantern \
    --check canvas brightspot=present in lantern \
    --check canvas brightness<=0.3 in sky
#
#  gate 'milestone-0.1': PASS
#    [ok] 'Wonder Book' reachable             measured=reachable_on_all_viewports
#    [ok] 'Main Menu' reachable-any           measured=reachable_on 2/3_viewports
#    [ok] perf fps>=55                        measured=58.2
#    [ok] canvas brightness>=0.6 in lantern   measured=0.71
#    [ok] canvas brightspot=present in lantern measured=present
#    [ok] canvas brightness<=0.3 in sky       measured=0.18
#  exit 0
#
#  When any check fails, exit 6 with the measured value AND the ROI on
#  the failed line. exit 5 if a --click step missed (silent skips are
#  gone). fail-closed per D-FAIL-CLOSED: if perf couldn't measure because
#  the page never settled, the gate says INCONCLUSIVE (page_never_settled),
#  not a silent null.

v0.3.3 receipts document the journey-on-Inspector heal plus WebGL canvas decode validated end-to-end (Pillow correctly reads gl.clearColor swapchain frames), NPR-shape fail-closed naming the inconclusive reason, the new 2-token --check grammar, per-finding screenshot_capped traceability, and two conflict-flag warnings. v0.3.1 calibration receipts: HEALS_v0_3_0_to_v0_3_1.md. Same canonical URL set across every version: inspector_run_v0_3_0_*.json through inspector_run_v0_3_3_*.json, no cherry-picked results.

What you actually use it for

Six shapes the same substrate solves. Drive-mode for agents and tests, audit-mode for reviewers.

How it actually works

Three primitives, one loop.