One tool. One endpoint.

Quickstart

Claude Code, one line. It registers a single tool, digest_url, and the agent calls it by itself whenever it meets a social link it cannot open.

claude mcp add --transport http linkdigest \
  https://linkdigest.dev/mcp \
  --header "Authorization: Bearer ld_live_..."

Cursor

Add this to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "linkdigest": {
      "url": "https://linkdigest.dev/mcp",
      "headers": { "Authorization": "Bearer ld_live_..." }
    }
  }
}

First call over REST

curl -X POST https://linkdigest.dev/api/v1/digest \
  -H "Authorization: Bearer ld_live_..." \
  -H "Content-Type: application/json" \
  -d '{"url": "https://www.youtube.com/watch?v=CEvIs9y1uog"}'

That link is a cached example, so it answers in about a second and costs nothing.

Authentication

Every request carries your key as a bearer token. Keys are created on the API keys page and shown once. A key can be given its own monthly credit cap, so one handed to a script cannot drain the account.

Authorization: Bearer ld_live_...

No key at all: the first look

A single link can be read with no key, no account and no card, and comes back as Markdown an agent can read straight away:

curl -sL https://linkdigest.dev/read/https://www.youtube.com/watch?v=CEvIs9y1uog

One uncached link per visitor; links already read are free as often as you like. YouTube is not included. A read still running answers 202 with a job id and a poll URL. The same call is available as /read/?u=<url-encoded link>.

Accepted links

Paste the link in whatever shape the app gave you. Share-button short links, full page addresses, a profile page with the post open, and the whole share-sheet text with the link buried in it all work — the first link is taken out for you. Measured 2026-09-13: 23 of 24 shapes read in full.

Discovery

Everything above starts from a link you already have. Discovery is the step before: from a topic to a set of links, so an agent can go topic → links → digests → one page on its own. There are two routes, and they are different because the platforms are.

From a search API

For xiaohongshu, douyin, youtube, x, web the MCP tool search_links and POST /api/v1/search find posts on a topic through a web search scoped to the platform, filtered to links a digest can read. A search that finds posts costs 4 credits on Douyin and Xiaohongshu and 1 elsewhere; one that finds nothing is free. Searching is metered by credits, not by an hourly count, and free accounts also share a pool of 30 searches a day; with digest: true the results become a batch with a comparison, billed as any batch is. For Xiaohongshu this returns the notes the web has indexed, and each is read from its public page — the page a logged-out phone gets — with no login, no session and no signature of ours. It is a subset of the platform: what search engines have seen, not your feed.

From a page you can see

For your own Xiaohongshu feed, or a search the web has not indexed, open the page in your own browser and press the browser extension’s Digest the notes on this page button (on Douyin, Digest this tab on a video): it starts a batch from the links on the tab you are looking at, labelled with the page, reading as many as you choose. The agent then calls list_batches (or GET /api/v1/digest/batches), takes the newest batch_id, and passes it to compare_links for the comparison. LinkDigest never holds your session. For Xiaohongshu every search answer also carries own_browser_url, the same search in your own browser; the batch the extension starts there is named after the query, so list_batches with label set to the query finds it. In the console this is the Search my Xiaohongshu button, which waits for that batch and shows it.

In the console

The same two routes without code: the Workspace’s Find posts mode runs the search and reads what you tick, its Many links → one page mode takes a pasted list (or the batch the extension started) and returns the comparison, and the Batches page lists every batch with its label — the console’s list_batches.

Research is those routes as a conversation, and it has its own place in the console: Research in the sidebar, also the fourth tab of the Workspace. Ask in plain words — 最近30天小红书上便携咖啡机哪个好 — and an assistant decides which platforms to search and which posts to read, says what it is doing as it goes, and answers in sentences with every point linked to a numbered source. 2 credits an answer; each search and each post read is charged as it is everywhere else, and a post already read is free. A follow-up carries what was read before. It is the console only; an agent does the same thing with search_links and digest.

Research skills put the job on the screen instead of a blank box: 开店前看同行, 定价, 选品, 竞品动作, 内容选题 and 游客口碑. One click runs a template — its own searches and platforms — and its answer must end with 「这周就做」: one to three steps for this week, each tied to the source number that justifies it. Same price as any other question.

Read one link. Returns the digest, or a job id when the media takes longer than a request can wait.

Example request

curl -X POST https://linkdigest.dev/api/v1/digest \
  -H "Authorization: Bearer ld_live_..." \
  -H "Content-Type: application/json" \
  -d '{"url": "https://www.xiaohongshu.com/discovery/item/6a6bf662000000002402de7e?xsec_token=CBCe01s1uyd2C8uMx6ycYcnEwPV5_31OX1u_I19FCSd54=&xsec_source=app_share", "translate_to": "en", "max_credits": 4}'

Example response

{
  "platform": "xiaohongshu",
  "author": "...",
  "title": "...",
  "posted_at": "2026-09-10",
  "caption": "...",
  "transcript": [{ "t": 0, "text": "..." }],
  "transcript_source": "asr",
  "ocr_text": ["...", "..."],
  "images": [{ "description": "...", "ocr": "..." }],
  "key_points": ["...", "..."],
  "key_point_support": [
    { "point_index": 0, "kind": "transcript", "t": 12.4, "image_index": null, "quote": "...", "verified": true }
  ],
  "coverage": {
    "seconds_transcribed": 95, "duration_seconds": 95, "largest_frame_gap_seconds": 4.2,
    "images_read": 1, "images_total": 1, "transcript_source": "asr",
    "not_captured": ["comments not read"], "links_found": 1
  },
  "links": [
    { "url": "https://arxiv.org/abs/1706.03762", "found_in": "on_screen", "label": "", "kind": "web" }
  ],
  "translation": { "language": "en", "key_points": ["..."] },
  "degraded": [],
  "cached": false,
  "credits": 13
}

Cached links, web pages and short posts come back as 200 straight away. A video that has to be downloaded and transcribed returns 202 with a job id. Collect it with the same key.

# 202 {"pending":true,"jobId":"abc…","poll":"/api/v1/digest/abc…","stage":"fetching"}

curl "https://linkdigest.dev/api/v1/digest/JOB_ID?wait=20" \
  -H "Authorization: Bearer ld_live_..."
# 202 while it runs, 200 with the digest when it is done.
# wait=0..20 holds the request until the job finishes (at most 20 s),
# so a loop with no sleep step (Coze, a Feishu shortcut) still works.

While it runs, stage says what is happening: resolving, fetching, transcribing 5:42 of audio · reading 42 frames, summarising, translating. Transcription and image reading run side by side.

Read a video you already have. Send its sound and a few frames; the video itself is never uploaded. This is the fallback for a Douyin video whose link the embed player will not serve, and for any video already on your machine.

The one command

The CLI extracts both pieces on your machine and makes the call. Python 3.11 or newer, and ffmpeg plus ffprobe on PATH.

pip install linkdigest-mcp   # needs ffmpeg + ffprobe on PATH

# Save the video first — in the Douyin app: 分享 → 保存本地
linkdigest capture ~/Downloads/video.mp4 --url "https://www.douyin.com/video/…"

# or, without a link:
LINKDIGEST_API_KEY=ld_live_... linkdigest capture ./video.mp4 --title "My video"

# see what would be sent, and charge nothing:
linkdigest capture ./video.mp4 --dry-run --verbose

A link passed in place of a file is refused with the save-then-rerun steps, never fetched. Markdown goes to stdout; --json gives the digest without raw_markdown, -o out.md writes it to a file. Exit codes: 0 ok, 2 the file or the account (unreadable, bad key, out of credits, rate-limited), 3 the server or the network.

Over HTTP

Multipart form data: one audio part, one frames part per JPEG, and one meta part of JSON.

ffmpeg -i video.mp4 -vn -ac 1 -ar 16000 -f mp3 -b:a 24k audio.mp3
ffmpeg -ss 4 -i video.mp4 -frames:v 1 -vf "scale='min(720,iw)':-2" -q:v 6 frame_00.jpg

curl -X POST https://linkdigest.dev/api/v1/capture \
  -H "Authorization: Bearer ld_live_..." \
  -F "audio=@audio.mp3;type=audio/mpeg" \
  -F "frames=@frame_00.jpg;type=image/jpeg" \
  -F 'meta={"page_url":"https://www.douyin.com/video/…","duration_seconds":65,"frame_times":[4]}'
# 202 {"id":"jobid.signature","poll":"/api/v1/capture/jobid.signature","retryAfter":3}

curl "https://linkdigest.dev/api/v1/capture/jobid.signature" -H "Authorization: Bearer ld_live_..."
# 202 while it runs, 200 {"status":"done","credits":13,"view_url":"…","digest":{…}}

A capture is private to the account that sent it: it is never cached under the post's key, never served to anyone else who asks for that link, and its id is signed with the account, so only that key can collect it. A video the caller already has is not a whole read of the platform and is not counted as one on platforms.

Everything else works as a link does: the same pricing by the second, the same digest fields, depth is always full. A refusal from the platform itself answers 503 with error_code platform_refused, how_to_read_now and a Retry-After header; nothing is charged.

LangChain and LlamaIndex

The Python SDK (linkdigest-mcp on PyPI, 0.2.5 or later) carries a LangChain tool and a LlamaIndex reader as optional extras; the base SDK still has no dependencies. The tool is named digest_url and passes LangChain's standard tool tests; the reader gives one Document per link with flat metadata (platform, title, author, source_url, credits).

pip install "linkdigest-mcp[langchain]"     # or [llamaindex]
export LINKDIGEST_API_KEY=ld_live_...

from linkdigest.integrations.langchain import LinkDigestTool
tool = LinkDigestTool(max_credits=20)          # the ceiling is yours, not the model's

from linkdigest.integrations.llamaindex import LinkDigestReader
docs = LinkDigestReader(depth="script").load_data(["https://v.douyin.com/xxxx/"])

Out of credits comes back to the model as a sentence with the pay link; a removed or private post as "cannot be read, nothing charged"; a long read as a job id it collects on the next call. The same package has linkdigest.links, an offline share-text parser (linkdigest-links "<share text>"). Source and examples: github.com/jcaiagent7143-ui/linkdigest-mcp.

Many links in one call: up to 10 on the free plan, 50 on Dev and 200 on Studio. Every link is billed exactly as a single call would be: cached links are free, credits are quoted before spend, the key’s cap applies. Collect by polling, or hand over a webhook and the summary is pushed to you.

curl -X POST https://linkdigest.dev/api/v1/digest/batch \
  -H "Authorization: Bearer ld_live_..." \
  -H "Content-Type: application/json" \
  -d '{"urls": ["https://www.youtube.com/watch?v=CEvIs9y1uog", "https://www.xiaohongshu.com/discovery/item/6a6bf662000000002402de7e?xsec_token=CBCe01s1uyd2C8uMx6ycYcnEwPV5_31OX1u_I19FCSd54=&xsec_source=app_share"],
       "webhook_url": "https://example.com/hooks/linkdigest",
       "translate_to": "en",
       "synthesis": true}'

# 202
{ "batch_id": "…", "status": "running", "total": 2, "done": 1,
  "failed": 0, "credits_total": 0, "poll": "/api/v1/digest/batch/…",
  "synthesis": null, "synthesis_credits": 0, "synthesis_error": null,
  "items": [ … ] }

A link that could not be read is an item with status: error and an error_code; the rest of the batch is unaffected. Batches are kept for seven days.

The synthesis object

With synthesis: true the batch stays running after the last link until the comparison is written, then carries it as synthesis:

Where the batch stands. Add ?include=digests to embed every finished digest, keyed by the url you sent.

curl "https://linkdigest.dev/api/v1/digest/batch/BATCH_ID?include=digests" \
  -H "Authorization: Bearer ld_live_..."
# 202 while links run, 200 when done

A competitor report

A batch sent with insight: true and synthesis: true is a competitor report: every note’s consumer insight plus one comparison across them. Over the API it is that batch, nothing new to call; in the console it is the Competitor report switch, and its batch page reads as a report to print, save as PDF or copy. It runs on credits from any plan: each note at its own price (typically 2) plus 1 for its insight, plus the comparison, so 3 notes come to about 10 credits. The console sizes it to the plan: 3 notes free, 20 on Dev, 50 on Studio.

This account’s last batches, newest first — including the ones the browser extension started. ?limit=1..10, default 10. One line per batch: batch_id, label, created_at, status, done / total / failed, credits_total and synthesis (ready, pending, failed, or null when none was asked for). Starts nothing, costs nothing.

curl "https://linkdigest.dev/api/v1/digest/batches?limit=3" \
  -H "Authorization: Bearer ld_live_..."

{ "batches": [
  { "batch_id": "…", "label": "护肤 早C晚A — 小红书搜索", "created_at": "2026-10-17T09:12:03.000Z",
    "status": "done", "total": 8, "done": 8, "failed": 0, "credits_total": 9, "synthesis": "ready" }
] }

Find posts on a topic on xiaohongshu, douyin, youtube, x, web. A search that finds posts costs 4 credits on douyin and xiaohongshu, 1 elsewhere, and credits says what it cost; one that finds nothing is free (402 when the balance is short). Searching is metered by credits, not by an hourly count. Add digest: true to read the results as one batch with a comparison in the same call.

curl -X POST https://linkdigest.dev/api/v1/search \
  -H "Authorization: Bearer ld_live_..." \
  -H "Content-Type: application/json" \
  -d '{"query": "retinol beginners", "platform": "youtube", "limit": 5, "digest": true}'

# 202
{ "query": "retinol beginners", "platform": "youtube", "credits": 1,
  "results": [ { "url": "https://www.youtube.com/watch?v=…", "title": "…", "snippet": "…" } ],
  "batch_id": "…", "batch": { "status": "running", "label": "retinol beginners", … } }

# platform: "douyin" (since 2026-10-21) → the video pages the web has indexed,
# newest first, each with the day it went up (read from its id)
{ "query": "露营装备", "platform": "douyin", "credits": 2,
  "results": [ { "url": "https://www.douyin.com/video/7611810685512633654", "title": "…", "snippet": "…", "posted_at": "2026-02-28" } ] }

Without a search key on the server the route answers 503 with error_code: search_unavailable; nothing else is affected.

The daily brief (每日稿, added 2026-10-31): watch a topic, and every day at hour_utc its hottest new posts on platform (xiaohongshu or douyin) are read (reads of them, 5 by default, up to 20 within the plan’s batch size), ranked and compared, and one post is written from them in draft_platform’s shape — in your voice when voice carries a sample — then mailed to the account’s address with a share link. Each run is the batch it starts: search + reads + comparison + draft, about 8–12 credits; a day with nothing new costs nothing, and an account that cannot pay is mailed instead and tried again the next day. The first run starts at once. Free plans watch 1 topic, Dev 5, Studio 20 (402 watch_limit past that). GET lists them; PATCH /api/v1/watch/{id} takes status (active / paused), run_now, voice and hour_utc; DELETE removes one. Nothing is published for you: the post is yours to paste, or to prefill with the extension.

curl -X POST https://linkdigest.dev/api/v1/watch \
  -H "Authorization: Bearer ld_live_..." -H "Content-Type: application/json" \
  -d '{"query": "露营装备", "platform": "xiaohongshu", "voice": "家人们，我是刚开始露营的小白……"}'

{ "watch_id": "…", "query": "露营装备", "platform": "xiaohongshu", "draft_platform": "xiaohongshu",
  "status": "active", "hour_utc": 1, "next_run_at": "…", "runs": 0, "last": null }

The Pain-Point Radar (痛点雷达, added 2026-11-07): one topic — a product, category or brand — and the run answers three things. What people search for after it: the suggestions Baidu, Taobao and Bing give (market cn) or Bing and YouTube (global), grouped by need. What the posts about it and their top comments complain about, want and ask: 12, 24 or 48 posts (size quick, standard, deep) from Xiaohongshu and Douyin (cn) or YouTube and TikTok (global), read with the consumer insight, every checked finding clustered into pain points, needs, purchase blockers, products, unanswered questions and gaps — each with count_posts, the number of posts it came from, counted by code, and quotes checked word for word. And, on a re-run of the same topic, what changed: new and rising points, new search suggestions. It answers 202 at once; poll GET /api/v1/radar/{id} (202 while it searches, reads and aggregates, 200 when done; ?format=md for the report as Markdown, and a share_url anyone can open for 30 days).

Billed as its parts: the two searches, each post at its own price with insight (+1), and the aggregation, 3 credits for 12 posts up to 6 for 48 — about 47 credits in all for a quick run, said as estimate before it starts, and fewer posts when the credits cover fewer. Fewer than 3 posts read: the aggregation is not charged. Without the credits for a minimal radar: 402 insufficient_credits with the pay links. weekly: true runs it again every week and mails what changed; PATCH turns that off. POST /api/v1/radar/preview answers the search suggestions alone, for 1 credit.

curl -X POST https://linkdigest.dev/api/v1/radar \
  -H "Authorization: Bearer ld_live_..." -H "Content-Type: application/json" \
  -d '{"topic": "扫地机器人", "market": "cn", "size": "quick"}'

{ "radar_id": "…", "status": "searching", "estimate": 47, "poll": "https://linkdigest.dev/api/v1/radar/…" }

curl https://linkdigest.dev/api/v1/radar/<radar_id> -H "Authorization: Bearer ld_live_..."

{ "status": "done", "posts_read": 12, "credits_total": …,
  "radar": { "sections": { "pain": [ { "label": "…", "count_posts": 5, "share": 0.42,
      "quotes": [ { "text": "…", "note": 3, "kind": "comment", "verified": true } ] } ], … } },
  "diff": { "since": null, "new_terms": [], … }, "share_url": "https://linkdigest.dev/r/…" }

What this key’s account has left: plan, rolling window, credits used and remaining, pack credits, the key’s own monthly cap, and its webhook_secret. Starts nothing, costs nothing. failed lists this key’s last 10 reads that did not go through, each with the answer that names its fix (a Xiaohongshu link needs its share token; a Douyin video its player will not serve can be uploaded). Failed reads cost nothing.

curl https://linkdigest.dev/api/v1/usage -H "Authorization: Bearer ld_live_..."

{ "plan": "dev", "unit": "credits",
  "period": { "window": "30d", "since": "…", "renewsAt": "…" },
  "used": 104, "limit": 500, "remaining": 646, "pack_credits": 250,
  "key": { "hint": "…", "monthly_cap": 30, "used": 13, "remaining": 17,
           "webhook_secret": "…" },
  "max_duration_seconds": 1800,
  "failed": [ { "at": "…", "url": "https://www.xiaohongshu.com/explore/…",
                "error_code": "unreadable", "message": "This Xiaohongshu link has no share token, …" } ] }

Public, no key. For each platform: its state, an ok / degraded / unknown status, and when a real post on it was last read in full.

Webhooks

The webhook receives the same summary as the poll, as JSON, with two headers. Only https targets on public hosts are accepted. Delivery is at-least-once: a non-2xx answer is retried after 1 m, 5 m, 30 m and 2 h, five attempts in all.

The secret is in GET /api/v1/usage under key.webhook_secret; it is derived from the key and retires with it.

Verifying a delivery

// Node
const [t, v1] = sig.split(",").map((p) => p.split("=")[1]);
const expect = crypto.createHmac("sha256", secret).update(`${t}.${rawBody}`).digest("hex");
crypto.timingSafeEqual(Buffer.from(expect), Buffer.from(v1));

# Python
t, v1 = (p.split("=", 1)[1] for p in sig.split(","))
expect = hmac.new(secret.encode(), f"{t}.{raw_body}".encode(), "sha256").hexdigest()
hmac.compare_digest(expect, v1)

Response fields

Status codes

Limits

What LinkDigest will not do is on the platforms page, in writing: X long-form articles expose only their lead image, an Instagram carousel gives only its first slide, and a Bilibili video the embed player will not serve is refused rather than guessed at. Anything that could not be read is named in degraded, and a digest that read nothing is not charged.

How each platform is opened — embed players, residential exits, what is stored and for how long — is on data sourcing & compliance.