What this downloader cannot do
Last reviewed 2026-11-04. Everything on this page was read from the code that serves the site, not estimated from how a downloader usually behaves.
Most pages about a TikTok downloader list what it does. This page lists what it refuses, what it answers when it refuses, and the three things that are genuinely absent from the download path: a rate limit, a result cache, and a size cap. The transcript tool is the one exception, and it carries two limits of its own further down.
Three endpoints, three different ways to fail
HTTP 200 does not mean the lookup worked. The resolve endpoint answers 200 even when every parsing source failed; the failure is inside the body as {"ok": false, "error": "..."}. The download endpoint is the opposite — it fails with a plain-text 400 or 502, with no JSON at all. A client that branch on the status line alone will treat failed lookups as successes, and a client that parses every error as JSON will crash on the download errors. Read ok for resolve, read the status code for download, and do not assume the same shape from both.
Which endpoint answers what, and in which format
| Endpoint | Success | Failure | Body on failure |
POST /api/resolve | 200, ok:true + data | 200, ok:false + error | JSON |
GET /api/download | 200, the video stream | 400 or 502 | Plain text |
POST /api/transcript | 200, ok:true + text and segments | 200 ok:false, or 429 when over the hourly budget | JSON |
GET /api/health | 200 always | — | JSON |
GET /api/stats | 200 | — | JSON |
| any unknown path | — | 404 | An HTML page |
There is no rate limit on downloads, and that cuts both ways
There is no rate limiter on the download path. Nothing there counts requests per address, nothing opens a window, nothing returns a 429. The transcript endpoint is the one place with a budget: one hour per address, five calls, then a 429 — the section below spells it out. If you are used to downloaders that allow twenty lookups a minute, the download endpoints have no such number; the only budget line on this page is the transcript one.
That is a description of the code, not an invitation. The machine reports 3.8 GiB of memory, and every lookup fans out to three upstream services with a 10-second timeout each. A script that opens fifty lookups at once will slow the site down for everyone including itself. Behave as if there were a limit of a few requests a second.
The transcript tool, and its two hard limits
Everything above describes the download endpoints. POST /api/transcript, which turns a video's speech into text, is the one part of the site with a real budget — it runs a Whisper model on the same machine that serves the downloads.
- 90 seconds of audio per request. The constant is
_ASR_MAX_SECONDS = 90; a longer post is cut at 90 seconds of audio.
- Five transcripts per visitor per hour, counted per IP address. The constant is
_ASR_PER_HOUR = 5. The sixth request inside that hour answers 429 with "转写请求太频繁,请一小时后再试", and a blocked attempt does not extend the window.
Two behaviours to know before relying on it. The model is base, kept small so a 60-second clip finishes in roughly half a minute on this hardware: accuracy varies with language and audio quality, and a music-only or silent clip comes back as {"ok": false, "error": "没有识别到语音(可能是纯音乐或无声视频)"} rather than an empty transcript. Transcription is also serialised behind a lock, one job at a time, so simultaneous requests queue instead of failing.
There is no result cache either
This is the opposite of what most tools in this category advertise. There is no stored copy of a previous answer, so asking for the same link twice performs the full upstream round trip twice. Two things follow, and both matter if you are writing something against this site:
- The second answer can differ from the first, because a source can change between the two calls. A result is a measurement, not a record.
- A repeated failure costs exactly as much as the first failure. Retrying a link that just failed does not get cheaper.
The 12 limits
Each row names the constant or the function behind it, so you can check rather than trust. Two of the twelve are the absence of a limit, which is stated plainly because competitors imply the opposite.
Every limit, with the code that decides it
| # | Item | Value | Where it comes from |
| 1 | Video lookups per address | no limit | No limiter guards the lookup path. The only counter in server.py is the transcript budget (five per hour per address, in the section below). |
| 2 | Downloads per address | no limit | Same reason. The download handler counts nothing — it opens the upstream connection as soon as the request arrives. |
| 3 | Lookup results cached | 0 — nothing is cached | There is no result cache. The same link twice means three upstream calls twice, not once. Opposite behaviour to downloaders that advertise a 30-minute cache. |
| 4 | Parsing sources tried | 3, in order | tikwm, then TikTok's own page JSON, then tikwm's mirror domain. The first one that returns a video address wins and the rest are never contacted. |
| 5 | Attempts per source | 1 | resolve_with_retry is called with retries=1, so the loop body runs once. The time.sleep(0.4) inside it is dead code and never executes. |
| 6 | Pause between sources | none | Because each source is tried once and the sleep is unreachable, three failing sources run as fast as they fail. A dead network fails in seconds. |
| 7 | Time to reach a source | 10 seconds | urlopen(..., timeout=10) in each of the three parsing functions. Short link expansion gets 20 s, and the resolve path hands it 10 s. |
| 8 | Download attempts per candidate | 3 | for attempt in range(3), with 0.6 s then 1.2 s of backoff. The comment above this code says two attempts — the code says three. Trust the code. |
| 9 | Download candidates | as many as you pass | The url parameter takes several addresses joined by a pipe. Each is tried up to 3 times, with a 0.5 s pause before moving to the next. |
| 10 | Time to reach the CDN | 30 seconds | urlopen(..., timeout=30) in the download path, per attempt. Three attempts across two dead candidates is the worst case a single request can take. |
| 11 | Maximum video size | none | Nothing compares a byte total anywhere. The file is streamed through in 256 KB chunks and written straight to the browser. |
| 12 | Maximum video duration | none | No duration check exists. A duration field is read from the source and passed through to the page, and that is all that happens to it. |
A comment that is wrong about its own code. The download handler says, in a comment, that each candidate is tried two times. The code underneath is for attempt in range(3), which is three times, with 0.6 s and then 1.2 s of backoff. This page reports three because three is what runs. It is mentioned here because a comment you can trust is worth more than a comment that sounds reassuring.
One more dead line. The retry helper contains a time.sleep(0.4) between attempts. It is called with retries=1, so the loop body runs exactly once and the sleep never executes. There is no pause between parsing sources in practice, which is why a lookup against a dead network fails quickly instead of crawling.
And one stale number. A comment in the stats module says this machine has 1.9 GB of memory. It has 3.8 GiB — the comment was true when it was written and nobody updated it. It is noted here only because it is the third place on this page where the code's own description of itself is out of date, and a reader deserves to know that the numbers here were measured rather than copied out of comments.
Every way a request can fail
This site does not use a list of named error codes. There is no enumeration to quote, so what follows is every distinct failure a caller can actually run into, with the status code and the body shape for each. The bodies differ, and that is the point.
Six failures, three body shapes
| What happened | HTTP | Endpoint | What the body looks like |
| not a TikTok link | 200 | resolve | The body is {"ok": false, "error": "..."} with an English-and-Chinese message. HTTP 200. Nothing is retried. |
| all three sources failed | 200 | resolve | Still HTTP 200. ok is false, error says every channel failed, and a detail array carries one line per source explaining why each one failed. This is the most useful field on the whole API and the easiest to miss. |
| unhandled exception | 500 | resolve | The only way this endpoint returns anything other than 200. The body is still JSON with ok:false. |
| no url parameter | 400 | download | Plain text: bad url. Not JSON. A parser expecting JSON throws here. |
| every candidate failed | 502 | download | Plain text: fetch failed: followed by the last upstream exception. Not JSON, and worth retrying later. |
| unknown path | 404 | any | An HTML page with links back to the site, not a JSON error. Scripts that assume every miss is JSON will break on this one. |
Retrying, in one sentence
Retry the download 502 — a CDN blip is worth a second attempt — and do not retry a 200 with ok:false unless minutes have passed, because the sources are the thing that failed and they do not recover inside a retry loop. The one useful exception is the detail array on a total failure: it names what each of the three sources said, and if all three say the same thing then retrying is pointless.
What it deliberately does not do
It does not remove a watermark the creator burnt in
The platform watermark that TikTok stamps on the file is stripped. A handle or logo the creator rendered into the video itself is part of the picture and stays there. No tool can remove that without re-encoding.
It does not upscale, and does not invent an HD version
The HD button points at whatever the uploader posted. A 480p original does not become 1080p here, and there is no transcoding step at all — the bytes that leave TikTok's CDN are the bytes you receive.
It does not sign in to TikTok
Every read is anonymous. Private, friends-only, age-restricted and deleted videos cannot be fetched, and that is the correct answer rather than a bug waiting to be fixed.
It does not convert audio itself
The MP3 you get is the audio track TikTok serves for that post. There is no ffmpeg on this server — /api/health reports ffmpeg as false — so no conversion, no re-encoding and no trimming happens here.
What is stored, and for how long
Three answers, one of which is not zero
| Item | What happens to it |
| The video | Never written to disk. It is relayed from TikTok's CDN to your browser in 256 KB chunks as it arrives, and dropped the moment the connection closes. |
| Lookup results | Never stored either. There is no cache layer, so a second lookup of the same link costs another round of upstream calls and can return a different result if a source changed in between. |
| One thing that is written | A counter file, stats.json, holding totals and per-day counts for visits, lookups and downloads. It stores no URLs, no video data and no addresses — the visitor tally keeps a list of addresses for the current day only, capped at 5,000 entries. |
Questions this page gets asked
Does a failed lookup return an HTTP error code?No. The resolve endpoint answers HTTP 200 for every failure it can describe: a link that is not a TikTok URL, and a lookup where all three sources failed. Only an unhandled exception produces anything else, and that is a 500. Read the ok field, not the status line.
Why does the download endpoint behave differently?Because it is a different kind of endpoint. Resolve returns a data structure, so failures are described inside that structure. Download streams a file, so there is no JSON to put a flag in — a failure has to be a status code. That is why a missing url is a plain-text 400 and a dead CDN is a plain-text 502.
Is there a rate limit?Not on downloads. The download endpoints count nothing, which is unusual for this category. The transcript endpoint is the exception: five transcripts per hour per address, and the sixth call in that hour answers 429. Beyond that, this is a statement about the code, not an invitation: the machine has 3.8 GiB of memory and every lookup fans out to three upstream services. Hammering it will make it slow for everyone.
Are lookups cached?No, and this is worth knowing before you build on it. There is no result cache, so asking for the same link twice performs the full upstream round trip twice. Two consequences: the second answer can differ from the first, and a repeated failure costs exactly as much as the first one.
What is the difference between the three sources?tikwm is a community API that has stayed up for years and returns the cleanest data. TikTok's own page JSON needs no third party but breaks whenever the page structure changes. The tikwm mirror is the same API on a second domain, used when the first one is unreachable. They are tried in that order and the first success ends the lookup.
Why does a link that worked a minute ago fail now?Because two of the three sources are outside this site's control. A source can be down, rate-limiting this server, or returning a page layout that no longer matches. There is no cache to fall back on, so a link is only as reliable as the sources are at that moment.
Is there a maximum video size?No. Nothing in the download path compares a byte total. The file is read in 256 KB chunks and each chunk is written to the browser immediately, so a 200 MB video takes the same code path as a 2 MB one.
Is there a maximum video length?No. There is no duration comparison anywhere in the server. A 40-minute video and a 4-second video follow identical code.
Can it download private or friends-only videos?No. Every read is anonymous and there is no login path, so a post that requires an account session is unreachable. Deleted and age-restricted posts fail the same way.
Does it re-encode the video or change the quality?Neither. The bytes that leave TikTok's CDN are the bytes your browser receives. No transcoding step exists here, which is also why the site reports ffmpeg as false — there is nothing to transcode with.
What does the demo mode do?When the page cannot reach TikTok it can fall back to a bundled sample clip so the interface still demonstrates itself. The response says so in a note field and the data is marked source: demo. Treat it as a UI sample, not as a real lookup — the id and title belong to the sample.
How can I check any of this myself?Send a profile link instead of a post link to the resolve endpoint. The status line will say 200 and ok will be false. Then request /api/download with no url parameter and look at the reply — it is plain text, not JSON. Those two requests tell you whether your client reads this API the way it is written.
Checking any of this yourself
Three requests settle it, and none of them needs an account.
- Send the resolve endpoint a profile link instead of a post link. The status line says
200 and ok says false. That is the whole point of this page in one request.
- Call the download endpoint with no
url parameter. The reply is the plain text bad url with a 400 — not JSON.
- Look up the same link twice. Because nothing is cached, both calls reach the sources, and the two answers can differ.
Paste a TikTok link
Free, no sign-up, no app. Public videos only.
Open ClipDocky