@browserless.io/mcp@1.14.0: Suspicious npm security report (Warn)

Loading npm security reports…

@browserless.io/mcp@1.14.0: Suspicious npm security report (Warn) | LPM Firewall

build/src/skills/sites/vaccines.gov/find-appointment-55bu87/SKILL.mdView file

Lines 58-98markdown

58The optimal path is **hybrid**: call the same Google Places `searchText` endpoint that vaccines.gov calls (one HTTP POST, no auth beyond a publicly-embedded API key, ~0.5s), filter results to the requested chains, then construct per-chain scheduler URLs. The browser path is a fallback when the API key has been rotated or rate-limited; it pays ~50× the cost and still returns the same data (vaccines.gov's frontend is a thin React wrapper over the exact API call below).

60### 1. Geocode the ZIP → lat/lng bounding box

62Vaccines.gov uses a `getBoundsOfDistance({latitude, longitude}, radiusMeters)` helper to convert a ZIP centroid + radius into a `{low, high}` rectangle. Any geocoding source works; the simplest is the Google Geocoding API or a static USPS ZCTA centroid table. Default radius in the vaccines.gov UI is **8047 meters (~5 miles)**; raise to **16093 m (~10 mi)** or **40234 m (~25 mi)** for sparse rural ZIPs.

64Example for ZIP 10001 (NYC, centroid 40.7506 / -74.0014, 5-mile radius):

66```json

67{

68 "low": { "latitude": 40.6783, "longitude": -74.0966 },

69 "high": { "latitude": 40.8228, "longitude": -73.9062 }

70}

71```

73### 2. Call Google Places `searchText`

75```

76POST https://places.googleapis.com/v1/places:searchText

77 Content-Type: application/json

78 X-Goog-Api-Key: AIzaSyCVvG6ZXIlxy76PCbqGnf9cP37XtA3HE-M

High

High Secret

Package contains a high-severity secret pattern.

build/src/skills/sites/vaccines.gov/find-appointment-55bu87/SKILL.mdView on unpkg · L78

High

Secret Pattern

Google API key in build/src/skills/sites/vaccines.gov/find-appointment-55bu87/SKILL.md

build/src/skills/sites/vaccines.gov/find-appointment-55bu87/SKILL.mdView on unpkg · L78

79 X-Goog-FieldMask: places.id,places.displayName.text,places.nationalPhoneNumber,

80 places.formattedAddress,places.addressComponents,places.location,

81 places.websiteUri,places.rating,places.primaryType,

82 places.regularOpeningHours.openNow,nextPageToken

84{

85 "textQuery": "pharmacy zip code 10001",

86 "languageCode": "en",

87 "includedType": "pharmacy",

88 "rankPreference": "RELEVANCE",

89 "strictTypeFiltering": true,

90 "regionCode": "US",

91 "pageSize": 20,

92 "locationRestriction": { "rectangle": { "low": {...}, "high": {...} } }

93}

94```

96The key, request shape, and field mask are exactly what vaccines.gov ships in `/_next/static/chunks/app/[locale]/results/page-*.js` (verified 2026-05-18). Paginate with `pageToken` from `nextPageToken` — vaccines.gov caps at **3 pages × 20 results = 50 max**. Iterate `pageSize=20` × up to `c<3` page tokens, accumulating until `o.length >= 50` (the literal cap from the bundle).

98**Transport note (Browserless):** This is a **cross-origin** call to

Lines 133-173markdown

133`browserless_agent` call carrying `proxy: { proxy: "residential" }` (the site sits behind

134Akamai — `akamai_visit_id` cookie is set on first response, and bare sessions get

135challenged). Keep all steps in **one** call's `commands` array so cookies/session persist:

136

1371. `{ "method": "goto", "params": { "url": "https://www.vaccines.gov/en/", "waitUntil": "load", "timeout": 45000 } }`

1382. `{ "method": "type", "params": { "selector": "<zip input>", "text": "10001" } }`, then `{ "method": "click", "params": { "selector": "<search button>" } }`

build/src/skills/sites/youtube.com/extract-transcript-loeude/SKILL.mdView file

Lines 247-287markdown

247 { "method": "waitForTimeout", "params": { "time": 3000 } }, // let the player chrome render

248 {

249 "method": "evaluate",

250 "params": {

251 "content": "(async () => { const pr = window.ytInitialPlayerResponse || null; if (!pr) return JSON.stringify({ error: 'no_player_response' }); const tracks = pr?.captions?.playerCaptionsTracklistRenderer?.captionTracks || []; let events = null; if (tracks.length) { const url = tracks[0].baseUrl + '&fmt=json3'; try { const j = await fetch(url).then(r => r.json()); events = j.events || null; } catch (e) { events = null; } } return JSON.stringify({ playabilityStatus: pr.playabilityStatus, videoDetails: pr.videoDetails, tracks: tracks.map(t => ({ baseUrl: t.baseUrl, languageCode: t.languag ...

252 },

253 },

254 ],

255}

256```

257

258The `evaluate` result comes back under `.value` — parse `playabilityStatus` / `videoDetails` / `tracks` exactly as in steps 3–5, and normalize the fetched `events[]` into segments per step 5. No session-release step is needed: there's nothing to release, and the session persists across separate calls (keyed by the call's `proxy`/`profile`). The read-and-fetch is batched into the one `commands` array above mainly because the signed `baseUrl`s are time-limited (~6 h) — fetch them promptly rather than deferring to a later call.

259

260A stealth + residential-proxy session is recommended for this fallback because YouTube's bot detection is more aggressive on the consent / `/watch` HTML path than on the InnerTube API — add `proxy: { proxy: "residential" }` (optionally `proxyCountry: "us"`) as a top-level arg on the call. But the API path itself in step 3 routinely succeeds from datacenter IPs with no proxy, so **don't pay for a residential proxy until you actually need it** — try the call without `proxy` first.

261

262## Site-Specific Gotchas

263

264- **`lengthSeconds` is a string**, not a number — JSON-parse coerces correctly but a naïve `videoDetails.lengthSeconds + 1` will concatenate. Cast.

265- **`captions` field is entirely absent**, not `null`, when the uploader has disabled captions. Distinguish `"captions" in player` vs `captions.playerCaptionsTracklistRenderer.captionTracks.length === 0` — both indicate "no transcript", but the former is the uploader's choice and the latter is occasionally a transient API state. Retry once on the empty-array case before declaring `captions_disabled`.

266- **Auto-generated detection: `kind === "asr"` is the canonical signal.** `vssId` starting with `a.` is a redundant cross-check. Don't try to infer auto-generated from text quality / lowercasing / no-punctuation — modern ASR adds capitalization and punctuation; that heuristic is dead.

267- **The InnerTube `key=` query parameter is no longer required for the `ANDROID` and `IOS` clients** as of late 2024 — those clients are validated by `User-Agent` + `clientVersion` instead. The `WEB` client still requires the key, which you harvest from `https://www.youtube.com/embed/<id>` HTML (`"INNERTUBE_API_KEY":"..."` — verified live as `AIzaSyAO_FJ2SlqU8Q4STEHLGCilw_Y9_11qcW8` on 2026-05-18, rotates ~quarterly). Don't hardcode the key.

High

Secret Pattern

Google API key in build/src/skills/sites/youtube.com/extract-transcript-loeude/SKILL.md

build/src/skills/sites/youtube.com/extract-transcript-loeude/SKILL.mdView on unpkg · L267

268- **`captionTracks[].baseUrl` is signed and time-limited.** The signature embedded in the URL expires after ~6 hours. Fetch the track within minutes of getting the player response; don't store baseUrls in a long-lived cache.

269- **`&fmt=json3` is mandatory for machine consumption.** Default response is `xml` (TTML-style) with HTML entities, font tags, and inline `<br>` — much harder to parse cleanly than json3's `events[].segs[].utf8`.

270- **Bare `https://www.youtube.com/api/timedtext?v=<id>&lang=<code>` GETs return HTTP 200 with empty body** when no signature is supplied. Don't be fooled by the 200 — the body length is 0. Verified 2026-05-18: `/api/timedtext?v=dQw4w9WgXcQ&lang=en&fmt=json3` → `200 OK Content-Length: 0`. The signed `baseUrl` from the player response is the only working entry point.

271- **`type=list` on the timedtext endpoint is deprecated** and also returns 200 + empty body. Use the InnerTube `/player` response's `captionTracks` array instead.

272- **Live streams may have no caption tracks even when `playabilityStatus === "OK"`.** Check `videoDetails.isLiveContent` and `videoDetails.isLive`; if either is true and `captions` is missing, report `live_stream_no_transcript` rather than `captions_disabled`. Once a live stream ends and is post-processed (typically within an hour), captions may appear.

273- **Shorts have transcripts.** A YouTube Short (`/shorts/<id>`) is just a regular video with portrait aspect ratio. The same InnerTube call works; the only difference is `lengthSeconds` is usually ≤60.

274- **`embedded_player_response` inside the embed page does NOT contain caption tracks** when fetched without a valid `Referer`. The embed HTML returns `previewPlayabilityStatus.errorCode: "PLAYABILITY_ERROR_CODE_EMBEDDER_IDENTITY_MISSING_REFERRER"` and the `captions` field is absent. This is a common dead-end. Always use the InnerTube POST instead. (Confirmed 2026-05-18 against `https://www.youtube.com/embed/dQw4w9WgXcQ` — 128 KB HTML, INNERTUBE_API_KEY and clientVersion present, but no `captionTracks` anywhere in the document.)

275- **The watch page is large.** `https://www.youtube.com/watch?v=<id>` consistently returns > 1 MB of HTML (verified — exceeded the Browserbase Fetch 1 MB cap on `www.youtube.com`, `m.youtube.com`, `music.youtube.com`, `/shorts/`, and `/watch_videos?video_ids=...` variants on 2026-05-18). Don't try to fetch and regex it from a lightweight fetch endpoint; either use the InnerTube POST or open it in a real browser session and read `window.ytInitialPlayerResponse`.

276- **`Origin: https://www.youtube.com` header on the InnerTube POST is recommended** even from the ANDROID client — it appeases the upstream WAF on rare 429-rate-limited paths. The `User-Agent` should match the `clientVersion`: `com.google.android.youtube/<version> (Linux; U; Android 14) gzip`.

277- **Region locks come back as `UNPLAYABLE`** with `reason: "Video unavailable in your country"`. The `ANDROID` client doesn't bypass these any more than the `WEB` client does — both honor geofencing. Use a residential proxy in the relevant region if you need to access region-locked content.

278- **Age-restricted videos return `LOGIN_REQUIRED`** on cookieless InnerTube. There's no clean public bypass; the legacy `EMBEDDED_PLAYER` cipher trick stopped working in 2023. Report `age_restricted` and move on, or fall back to a logged-in browser session if the caller has cookies.

279- **Caption tracks may be empty arrays even on healthy videos.** Some videos have `captions.playerCaptionsTracklistRenderer.audioTracks` populated but `captionTracks: []` — these are videos with multi-language _audio_ dubs but no subtitle tracks. Treat as `captions_disabled`.

280- **Translation tracks via `&tlang=` are machine-translated by Google.** They're not separate tracks in `captionTracks`; they're a per-baseUrl query parameter. Available target languages are listed in `captions.playerCaptionsTracklistRenderer.translationLanguages[]`.

281- **Multiple `segs[]` per event** on auto-generated tracks represent word-level timing for highlighting; concatenate them to get the line text. On human-authored tracks, multi-`segs` usually represents inline formatting (italics, color). Either way, concatenate `utf8` fields and you get the human-readable line.

282- **Empty `segs` events with `aAppend: 1`** are continuation markers for the previous event's last segment (used to extend the highlight window). Skip them — their text was already emitted.

283

284## Expected Output

285

286Six distinct outcome shapes. Always include the `video_id` and any metadata you successfully resolved, even on failure.

287

Long lines were clipped for display.

build/src/skills/sites/rec.us/sf-court-booking-c8l5z9/SKILL.mdView file

Lines 198-238markdown

198 "end_time_local": "13:30",

199 "duration_minutes": 90,

200 "price_cents": 750,

201 "currency": "USD"

202}

203```

204

205If during skill execution you cannot find a confirmation code on the post-book page, fall back to navigating to `https://www.rec.us/account/reservations` and reading the most-recently-created entry; its `id` UUID prefix is the canonical confirmation code.

206

207### 7. Session teardown

208

209No explicit release step — there is nothing to release. The session persists across calls keyed by `profile`/`proxy`. Batching the whole warm-up → date-pick → slot → login → book flow into the single `commands` array of step 5 keeps the Firebase session cookies together through to the booking POST and saves round-trips. If you do split the flow, repeat the same `profile`/`proxy` on each call to reconnect to the same logged-in session; a call that drops or changes it lands in a different, logged-out session.

210

211## Site-Specific Gotchas

212

213- **The API at `api.rec.us/v1` is public and unauthenticated for read endpoints.** No CORS shenanigans, no rate-limiting observed during skill generation, no captcha. `availability`, `schedule`, `price`, `sports`, `activities`, `organizations/{id}`, `locations/{id}` all return live JSON to anonymous callers. Use it. The Next.js front-end is a thin shell over this API.

214- **The site is not bot-protected.** No Akamai, no Cloudflare turnstile, no PerimeterX. A plain `browserless_agent` call (no `proxy` arg) loads the org listing fully hydrated; stealth is not needed and a residential proxy buys nothing but geo-stability. Skip the `proxy` arg entirely for this site.

215- **Login is Firebase Identity Toolkit, not a rec.us-owned endpoint.** The auth POST is:

216 ```

217 POST https://identitytoolkit.googleapis.com/v1/accounts:signInWithPassword

218 ?key=AIzaSyCp6DCwnx-6GwkMyI2G1b8ixYs4AXZc-7s

High

Secret Pattern

Google API key in build/src/skills/sites/rec.us/sf-court-booking-c8l5z9/SKILL.md

build/src/skills/sites/rec.us/sf-court-booking-c8l5z9/SKILL.mdView on unpkg · L218

219 Body: {"returnSecureToken": true, "email": "...", "password": "...",

220 "clientType": "CLIENT_TYPE_WEB"}

221 ```

222 Failure response is HTTP 400 `{"error":{"code":400,"message":"INVALID_LOGIN_CREDENTIALS"}}`. The Firebase Web API key in the URL is a public client identifier (not a secret) and was stable as of skill generation — pin it but treat key changes as a low-priority maintenance item, not a credential leak.

223- **Clicking the parent time-slot button does NOT reliably open the booking modal.** The button has two children: a "{H}:MM" paragraph and a div of duration cells ("60", "90"). Click the **duration cell** (`StaticText: "60"` or `"90"`), not the outer button. Verified: clicking the outer button left the page unchanged; clicking the inner duration text opened the modal correctly.

224- **Click "Log in" in the page header BEFORE clicking "Book" in the modal.** If you click Book unauthenticated, rec.us opens a _signup_ modal (with email, confirm-email, first/last name, phone, password, address fields). There is a "Already have an account? Log In" link to switch — but pre-authing avoids the back-and-forth entirely.

225- **Reservation windows differ per court at the same location.** At Alice Marble, Courts 1–3 are reservable 7 days in advance with 90-min fixed slots starting on 90-min boundaries (7:30, 9:00, 10:30, 12:00, 1:30, 3:00, 4:30); Court 4 is reservable only **2 days** in advance with 60-min slots on the hour. The `defaultReservationWindowDays` field on each court in the availability response tells you which.

226- **Reservations open at a fixed local time, not 24h ahead.** `reservationReleaseTimeLocal` defaults to `08:00:00` (Court 4 at Alice Marble: `12:00:00`). Trying to book outside the release window returns an error from the booking POST — there's no client-side guard. For the skill, check that `(date - today_local) ≤ defaultReservationWindowDays` AND that the current local time is past `reservationReleaseTimeLocal` if booking the maximum-window day.

227- **Booking limits are enforced server-side, not surfaced in the UI.** Alice Marble's `playGuidelines` field on `/v1/locations/{id}?publishedSites=true` says "may not book more than 1 court in a day or 3 courts in a calendar week" — but the booking modal still shows a "Book" button to users who would exceed it. The POST will fail with a (currently unverified) error response; document this as `reason: "booking_limit_exceeded"` in your output.

228- **Schedule slot keys use `"HH:MM, HH:MM"` with a comma-space separator**, and the `dates` map key is `"YYYYMMDD"` (no separator). Don't confuse with ISO format. Times are local to the location's `timezone` (`America/Los_Angeles` for all SF locations).

229- **`RESERVABLE` windows are contiguous, not pre-split.** A `"12:00, 16:30": {"referenceType":"RESERVABLE"}` entry means 4.5h of contiguous reservability, not a single slot. The booking-policy on the court (in `availability.courts[].config.bookingPolicies`) specifies fixed-slots — e.g. `{startTimeLocal:"12:00:00", endTimeLocal:"13:00:00"}` × N — that you intersect with the contiguous block to enumerate sub-slots. The UI does this client-side.

230- **Court selector defaults to "Court 1" but you must verify the chosen court is actually `RESERVABLE` at the chosen time** — the modal's court dropdown lists _all_ courts at the location, not just available ones. Pick the court from your step 4 enumeration, then switch the modal's combobox if it didn't preselect correctly.

231- **Pickleball + Tennis can share a court.** At Buena Vista, the location-availability shows the same physical court with `sports: [{name:"Tennis"},{name:"Pickleball"}]`. The schedule's `RESERVABLE` block is shared — a tennis reservation at 12:00 blocks pickleball at 12:00 on that court. Filter by `sport.name` _and_ by the schedule.

232- **URL filter param**: `?activityId={uuid}` on the org landing page filters the listing client-side. Useful for screenshots but not strictly required by the API path — the schedule API already returns per-sport data.

233- **Stripe checkout is in-page, not redirected.** A card-on-file (added during account creation) is used silently; no Stripe.js redirect to checkout.stripe.com observed. If the user has no card, expect the modal to show a "Add payment method" step before the final Book button.

234- **Per-location `noReservationText: "Not Reservable"` means walk-up-only courts.** A few locations (e.g. some Upper Noe slots) show "No free spots available - Check back soon" indefinitely — that's the location's status, not a transient stockout. Surface as `reason: "location_not_reservable"`.

235- **The org's `/discovery/programmed` endpoint returned `[]` during testing** — it's intended for programs/classes, not court reservations. Don't waste time on it.

236- **`_next/data/{buildId}/...` SSR data endpoints return only `{pageProps: {selectedTabId: null}}` for location pages** — the page does all real fetching client-side via `api.rec.us`. Don't try to scrape Next.js data routes for availability.

237- **Confirmation code format is not verified end-to-end.** The skill-generation run did not have a real account, so the post-booking confirmation page was not captured. The format documented above (8 hex chars matching the reservation UUID prefix) is inferred from `referenceId` UUID conventions visible in the schedule API. A future agent with a real account should verify and update this skill.

238

@browserless.io/mcp@1.14.0

Static Scan Results

Decision evidence

Source & flagged code

Source & flagged code

Findings

Source & flagged code