backlink-guard

Find the pages your searchers already land on, the sites that link to your competitors and not to you, the sites Google's AI Overviews cite for your keywords, and what AI assistants answer when asked for what you make; check a link source before you submit your site to it, check the link after it goes live, look at your own backlink profile for the patterns Google calls link spam, and find the links that point at pages of yours that are gone or redirect badly.

Most backlink tools help you send more. This one tells you where not to send, and when to slow down.

[How backlink-guard works: find the pages your searchers land on, and with competitors, cited and asked the sites your competitors earned and the sites AI answers cite, with --early the links they got at launch; vet each site (its own pages, GoAnyAPI, Glasser or AIsa, and with --newcomers whether it still takes new sites), get avoid, caution or ok with the evidence, follow the rules, submit, then verify and log; audit your own profile now and then, with the sites one owner runs; lost lists backlinks landing on 404 pages or redirects, with a suggested 301. What 0.11 added is orange]

Quick start

uvx backlink-guard vet mcpservers.org turbo0.com navi.tools producthunt.com clawhomepage.com

uv's uvx runs it straight from PyPI, with nothing to install; uvx backlink-guard@latest … picks up a new release. To keep a backlink-guard command installed: pip install backlink-guard (or uv tool install backlink-guard). Python 3.10 or newer, no dependencies. With no key it reads the sites' own pages, for free (--free asks for just that). The keys (GoAnyAPI, and Glasser or AIsa) add rating, visits, Google's index and spam scores: about $0.17 for one site and $0.46 for thirty, so check the sites you are considering in one run.

Checked with both keys on 2026-10-13 and 14:

Action is what to do with the site: skip, look first, free only, low value or submit. It reads three answers that are also shown apart. Link policy: what a link there would mean under Google's link spam policies, one finding being enough; mcpservers.org sells followed links, so list there for free and never pay. Value: what a listing can bring, which is the site's reach, or none when nobody can reach it (dead, parked, dropped by Google, a shortener, a network page). The verdict weighs every finding: one avoid finding makes avoid, and caution takes two caution findings. Risk is 3 for each avoid finding and 1 for each caution. A site its own pages put at avoid, like navi.tools, gets no paid check.

Every finding comes with the text or the numbers it was read from:

turbo0.com — ⚠️ A high rating with almost no visitors: the rating comes from links, not readers. rating 81, about 2,218 visits a month. ⚠️ It asks for a link or badge back as a condition. "Free $0 … Backlink to our site is required." (Google: Excessive link exchanges, or requiring a link as part of an arrangement)

Sites that share a page template or an analytics or ads account, with each other or with sites from an earlier check on the same machine, are flagged as one operator's. When most of them have something else against them and the shared spam-link list lacks them, the report ends with a link to a filled-in report for that list; nothing is sent until you submit it.

How well it works

The rules were chosen on 100 sites they had never seen (2026-10-16): 40 that Ahrefs marks SPAM and 60 it does not.

• Every normal-labelled site called avoid does not load or has no pages in Google.
• Of 16 sites around shared checkout links, all 10 network or hacked sites come out skip, and 5 of the 6 that merely cite those products come out submit.
• What a listing brings follows reach: on two sites' analytics, 2 of 37 and 1 of 73 referring domains rated low had sent a visit, against 10 of 14 and 2 of 7 rated high.
• The ten well-known directories and communities all come out submit.

Nine tests, with their data and what each changed: docs/validation.md.

Where to be: find

A link is first a way in for visitors. A page that ranks for your keyword already gets the people looking for what you make, so a mention there is the most relevant link there is.

python -m backlink_guard find "claude skills" "claude code skills" "agent skills" --site agentskillshub.top

Three keywords, 30 results each: $0.027 on 2026-10-14. Visitors are monthly searches times the share of clicks a position usually gets, a size and not a count. --vet then vets the directories and article sites it found; pages on GitHub, Medium, YouTube or Reddit are judged by the page and its author, not the platform.

Links your competitors earned: competitors

The sites that already write about your competitors are the ones to tell about what you make. For each competitor, competitors lists the referring domains first seen in the last 7 days (the ones already linking to you left out), then the sites that link to two or more of the competitors and not to you, the link intersect. Each comes with a page that links, what kind of page it is and what can honestly be done there.

backlink-guard competitors yoursite.example --vs competitor-a.example competitor-b.example competitor-c.example

competitor-a.example: 941 new referring domains; 1 already link to you, left out

Site	Rank	First seen	Linking page	Kind
agenthub.example	195	2026-10-19	Skill install and guide	article
toolsdir.example	23	2026-10-16	Agent Skills Marketplace	directory

Sites linking to two or more of them, not to you

Site	Links to	Rank	First seen	Linking page	Kind
compare.example	competitor-a.example, competitor-b.example	25	2026-10-09	Best Claude skills marketplaces compared	list
indies.example	competitor-b.example, competitor-c.example	51	2026-05-18	Competitor B, made by indies	directory

Left out, and named at the end: the sites on the shared spam-link list, the links that score 50 or more for spam (every row carries the score), and three or more sites whose linking pages share one title, one network's template, with a filled-in report for the list. On 2026-10-20, 86 of one competitor's 100 newest referring domains were such pages. Rank is what the site's links pass to the competitor, on the backlink index's 0 to 1,000 scale; first seen is when its crawler found the link.

The intersect is one call: the domain intersection in its partial mode returns the domains linking to at least two of the sites given (its page says all of them; tried live, it is two or more). About $0.025 a call, 2N + 2 calls for N competitors: three competitors cost $0.20 on 2026-10-20. --max-usd caps a run ($0.30 unless set): a call is made only while its estimate still fits, answers cached in the last 7 days are free, and the report names what the cap left out. Vet a site before you spend time on it; never pay for a followed link.

Their earliest links: competitors --early

Sorted by when the backlink index first found them, a competitor's referring domains begin with what it got in its first weeks: the directories, launch sites, lists and communities that took it when it was new, and take a new site still. The strong sites that link to it today mostly do not. --early reads that end of the list, for each competitor and for your own site:

backlink-guard competitors yoursite.example --vs competitor-a.example competitor-b.example --early

Site	Launched	30 days	60 days	90 days	Now, all
yoursite.example (you)	2026-03-08 (first archived copy)	4	13	19	165
competitor-a.example	2026-02-17 (first archived copy)	14	21	26	75

Launch checklist: in the first 90 days of two or more of them, not linking to you

Site	Of	First seen	Rank	Linking page	Kind	Took a new site	Free vet
launchdir.example	competitor-a.example, competitor-b.example	2026-05-06	0	Competitor A | LaunchDir	directory	✅ competitor-b.example at 31 days (2026-05-06)	🆓 free only; 2/7 new

• Launched: the Wayback Machine's first copy, or the first referring domain when that came first, or more than half a year later (the domain sat unused before).
• The same age: each site's referring domains 30, 60 and 90 days after its launch, lost ones counted, so you see the pace a site of your kind gets, not the profile it has built since. A week that brought five or more and three times its usual is named with its sites: launch week, as a rule, and the sites in it are the channels it used.
• Kept: still live, followed and not site-wide. One site linking from its footer or navigation, or from 100 pages or more, is mostly the owner's other sites or a partner, not a channel: those are named apart, and the lost and nofollow-only ones counted. The spam-list, spam-score and template rules above apply.
• The checklist: what two or more of them got early and you lack (with one competitor, its own list), each with a linking page. Took a new site: the youngest competitor it linked to, at that age; ✅ when that was in the last half year, so it still takes new sites. The first 10 get a free vet with --newcomers, which takes a minute or two (--vet 0 skips it).
• Cost: one call a site, yours included, for its earliest 1,000 referring domains (at most $0.06; $0.03 for a site with 165), and one call a competitor for linking pages: one competitor and you cost about $0.08 on 2026-10-27, three about $0.35 at most. --max-usd is $0.50 here unless set.

The backlink index begins on 2019-01-15: for a competitor older than that, its earliest links there are the index's first crawl, and the report says so. Compare against competitors that launched near your own launch.

Where AI answers come from: cited

An AI Overview cites the pages it was written from. For each keyword, cited reads Google's first page with its AI Overview (US, English), says whether the overview cites your site and where you rank, and then counts the sites cited across all your keywords: the places where a mention gets you into the answer.

backlink-guard cited yoursite.example "claude skills" "agent skills" "claude code skills" "agent skills hub"

$0.002 a keyword on 2026-10-20, $0.004 when the overview loads after the page; --max-usd caps it ($0.10 unless set). A keyword with no AI Overview is said so, which is not the same as not being cited: Google shows an overview when it judges one helpful, and another day, place or searcher can differ.

What AI assistants answer: questions and asked

People also ask ChatGPT, Perplexity, Claude and Gemini directly, and each of them searches the web before it answers. questions writes the questions to ask; asked puts them to five assistants and keeps what each one searched for, read and cited.

backlink-guard questions "a directory of agent skills" --plain > questions.txtbacklink-guard asked yoursite.example --file questions.txt --name "Your Brand" --rank

• questions: ChatGPT, Claude and Gemini each write the questions they expect to be asked (no web search), merged once: about $0.01.
• asked: each question to ChatGPT (its own interface), Google's AI Mode, Perplexity, Claude and Gemini (with their web search). For each assistant, how many answers wrote your domain, gave only your name, cited you, or named a site under your name with another ending (whose mentions are not counted as yours); then the sites and pages cited most, each with its kind and what can honestly be done there, and the searches the assistants ran (AI Mode shows none).
• --rank: each of those searches on Google and Bing (US, English, top 20). An assistant reads what its search returns, so a site missing from the results is rarely read at all, and one that ranks without being cited was passed over. The report says where you rank, and puts the sites on the first page of the most searches next to the answers that cited them.

On 2026-10-26, for agentskillshub.top: 8 questions, 56 answers, 67 searches. No answer read or cited the site; it was in Google's top 20 for 3 of the 57 searches Google answered and in Bing's for none of 67. The work is in the search results first.

ChatGPT and Perplexity are asked twice, since their answers vary most: about $0.12 a question with the default samples, $0.0035 a search on each engine for --rank. --max-usd caps the whole run ($1.00 unless set). Answers asked in the last 7 days come from the cache, so --rank after a run costs only the lookups.

Why

Google's spam updates came in March, June, August and September 2026. "Buying or selling links", "excessive link exchanges", "low-quality directory or bookmark site links" and "using automated programs or services to create links to your site" are all on Google's list of link spam. Directory lists passed around for link building go stale fast and say nothing about any of this.

A verdict here is a set of signals a person can check. It is not Google's judgment: nobody outside Google can tell whether one link is counted, ignored or held against a site. Google ignores most junk links. What the tool is good for is not wasting effort on dead, empty or link-selling sites, and not building the pattern yourself.

What it checks

Glasser or AIsa means whichever key is set, Glasser when both are: they resell the same endpoints at the same prices, but for two searches, find's ($0.006 a keyword through Glasser, $0.005 through AIsa) and cited's (up to $0.006 against $0.004). What Glasser does not sell stays on AIsa: the site: search that stands in when GoAnyAPI's fails, --deep's Ahrefs rating and Semrush numbers, and everything questions and asked ask. See Keys.

For your own domain (audit): totals and how many referring domains pass ranking credit (GoAnyAPI), new and lost referring domains by week, up to 1,000 of your backlinks one by one with the referring domains behind each anchor, the site-wide links (Glasser or AIsa), the spam score of the sites linking to you, and the referring domains one owner runs (their home pages, free). lost reads the same backlinks for the pages of yours they point at that are gone or redirect.

Your own links: audit and lost

A count of links says little on its own: a site that puts your link in its footer gives you as many links as it has pages. So every count of links comes with the number of referring domains (the registrable domain: blog.a.example and a.example are one, alice.github.io and bob.github.io two). One site linking to one page of yours from 20 or more of its pages (--sitewide changes it), or from three or more with most of them in its footer, sidebar or navigation, is a site-wide link: one source wherever a share is worked out, and the report says so.

backlink-guard audit yoursite.example

• Read one by one: 60 of 1,840 links from 21 referring domains (AIsa); 40 links from 1 site, counted once

ℹ️ Some sites link to you from many of their pages (site-wide): each counts as one source. 40 links from 40 pages of news.example to /blog/old-launch-notes, 100% in its footer, sidebar or navigation. Each site counted once in the shares here (Google: Widely distributed links in the footers or templates of various sites)

audit also reads the home page of each referring domain (free, at most 300, --homes; 0 skips it) for the analytics and ads accounts it reports to, as vet does: AdSense, Google Analytics, Tag Manager, Clarity, the Facebook pixel, Cloudflare Web Analytics. Three or more sites on one account, two of them linking to you, are one owner's network: its links are one owner's votes (count it as one source), and a network built to link is the pattern Google's spam policies describe. The accounts join the networks seen before, so a site you vet later is matched against them, and when the shared list already holds half of a network, the report links to a filled-in report for the rest. When the backlinks read name fewer sites than --homes (one site's strongest 1,000 backlinks came from 29 referring domains), the referring-domains list (Glasser or AIsa) fills up the rest, about $0.035 for 300.

⚠️ Several sites linking to you report to one analytics or ads account: one owner's network, whose links count as one source. 24 referring domains in 1 network: cloudflare:0a1b2c3d… on poker-casino.example, crime-news.example, mirror.example …, 26 links. Count each network as one source (Google: Using automated programs or services to create links to your site)

A media group's titles or a platform's hosted sites can share an account honestly: read the sites before you act. Google ignores most such links; a disavow file is for a manual action.

lost asks each page of yours that gets links whether it is still there: one GET each, redirects followed, four at a time, at most 200 pages, the most linked first. Only 404 and 410 make a page dead, also at the end of a redirect chain; any other error is reported apart, as refused or unanswered, never as dead. For each dead page it suggests where to send it: the live page in your own sitemaps (robots.txt or /sitemap.xml, indexes followed, up to 10 files and 50,000 pages) whose path is closest, with the match from 0 to 1, and no suggestion under 0.50.

backlink-guard lost yoursite.example

ℹ️ Sites link to pages of yours that no longer exist: the links reach nothing. 3 pages of yours answer 404 or 410 and have links from 12 referring domains … (Google: a page that has moved for good takes a permanent server-side redirect, 301 or 308)

Sorted by referring domains, a site-wide source counted once in Links. Nothing is redirected: it is a list to read, and a page with no good match may be better left a 404 or sent to a parent page you choose. Google's page on redirects is the reference.

The pages of yours that redirect are followed one hop at a time and judged, worst first: broken (a loop, more than ten hops, or no page at the end), soft-lost (it ends on your home page, on another site, or on a page whose path matches the old one under 0.50; Google's page on site moves says redirecting old URLs to one irrelevant page such as the home page "might be treated as a soft 404 error", so the closest live page is suggested as for a dead one), temporary (a 302, 303 or 307 on the way: Google keeps showing the old URL), chain (two hops or more: Google advises one) or fine.

Redirected well (one permanent hop to a page like it): 1 page(s), e.g. /zh/weekly → /zh/go-global/weekly.

Prices are what the calls cost on 2026-10-13 and 14. Paid answers are cached for 7 days.

A site its own pages already put at avoid (dead, parked, a URL shortener, a network template page, a link seller, links out to gambling sites) gets no paid check: the rest could only add to it. A site that does not load gets one question, its visits, since a site people visit that refuses the check is blocked, not dead. An ok whose check failed where it could have found something (rating, visits, index, backlink profile, checkout links, spam score, search history) is marked incomplete: run it again before relying on it.

Commands

backlink-guard find "<keyword>" [more ...] [--site yoursite.com] [--skip competitors.txt] [--vet] [--json]backlink-guard vet <site> [more ...] [--from list.xlsx] [--skip yoursite.com] [--limit 30] [--topic "keyword"] [--referrals ga4.csv] [--deep] [--recent] [--newcomers] [--free] [--fresh] [--json]backlink-guard competitors <your-domain> --vs <competitor> [more ...] [--days 7] [--limit 100] [--max-usd 0.30] [--json]backlink-guard competitors <your-domain> --vs <competitor> [more ...] --early [--vet 10] [--max-usd 0.50] [--json]backlink-guard cited <your-domain> "<keyword>" [more ...] [--max-usd 0.10] [--json]backlink-guard questions "<what you make>" [--count 10] [--plain] [--json]backlink-guard asked <your-domain> ["<question>" ...] [--file questions.txt] [--name "Your Brand"] [--engines chatgpt,ai_mode,perplexity,claude,gemini] [--samples N] [--rank] [--max-usd 1.00] [--json]backlink-guard audit <your-domain> [--deep] [--sitewide 20] [--rows 1000] [--homes 300] [--json]backlink-guard lost <your-domain> [--sitewide 20] [--rows 1000] [--json]backlink-guard verify <page-url> --to <your-domain>backlink-guard log add <site> --kind directory --anchor "Your Brand"backlink-guard logbacklink-guard recheck [--to <your-domain>] [--deep] [--free] [--json]

From a clone, python -m backlink_guard runs the same. Your log, the paid answers (kept 7 days) and the networks seen before are kept in ~/.backlink-guard, or in ./state when it is run from a clone that has one; BACKLINK_GUARD_STATE moves them.

• find: $0.005 a keyword and $0.012 for the search volumes; --vet adds the vet cost for the sites found. --site leaves your own site out, --skip competitors.
• vet: about $0.01 a site, plus about $0.16 for the batch (spam scores and search history), the same for one site or thirty: check the sites you are considering in one run; --topic adds a column with where each site ranks for your keyword ($0.005 a keyword); --free reads the sites only. --from reads the sites out of a list (text, Markdown, CSV, JSON or an Excel sheet, as paid backlink lists usually come); --skip leaves out your own site, competitors or what you already have (sites or a list file); --limit caps a run at 30 sites unless raised. --deep adds about $0.19 a site: the site's own backlink profile ($0.024), its technology ($0.012), how often ChatGPT's answers mention it ($0.10) and its checkout links on inner pages (about $0.05, plus $0.02 a link found); with an AIsa key also Ahrefs' rating ($0.08) and Semrush's numbers ($0.001); --recent asks whether Google is taking in its new pages ($0.008 a site, and slow: GoAnyAPI's search calls take 10 to 120 seconds). Use both before paying for a placement. --referrals reads your analytics export (GA4's Traffic acquisition by Session source, as CSV) and adds what each site has sent you. A check that finds one operator's sites missing from the shared spam-link list ends with a link to a filled-in report for it.
• competitors: about $0.025 a call, 2N + 2 calls for N competitors (three: about $0.20). --days sets what new means (7), --limit the referring domains read from each list (100, at most 1,000), --max-usd the cap on a run ($0.30); cached answers are free and what the cap left out is named. --early: about $0.06 a site, yours included, plus a call a competitor for linking pages (one competitor and you: about $0.08); --vet vets that many checklist sites free (10, 0 skips it); --max-usd is $0.50 unless set.
• cited: $0.002 a keyword through AIsa, $0.004 when the overview loads late; through Glasser $0.004 and $0.006; --max-usd caps it ($0.10).
• questions: about $0.01 for the three writers; --count sets how many each writes (10), --plain prints the list for a file.
• asked: ChatGPT and AI Mode $0.004 an answer, Perplexity about $0.006, Claude and Gemini about $0.045 (2026-10-21): about $0.12 a question, with ChatGPT and Perplexity asked twice. --samples asks every assistant that many times, --engines picks them, --name adds a name the answers may call you by. --rank adds $0.0035 a search on Google and on Bing (2026-10-26). --max-usd caps the run ($1.00): the most-run searches are looked up first, and the report names what the cap left out.
• audit: about $0.12, with --deep (full anchor distribution) about $0.16. Of that, about $0.06 is the 1,000 backlinks read one by one (--rows asks for fewer: 100 cost $0.028 on 2026-10-18). Without a Glasser or AIsa key the 20 top linking pages from GoAnyAPI stand in. The home pages of up to 300 referring domains are read for shared accounts (free; --homes changes it, 0 skips it), with about $0.035 more when the backlinks read name fewer sites than that.
• lost: the same backlinks, so nothing more after an audit of the same site in the last 7 days, else about $0.06; asking your pages (and following their redirects hop by hop) and reading your sitemaps is free.
• verify: free. Verdicts are followed link, qualified link (nofollow, sponsored or ugc), redirect only (through the site's own /go/ page, also when the target is base64-encoded), not found, unreadable. Only those three rel values hold back credit; noopener and external do not. Each link also says where it sits, read from the elements around it: body text, main area, header, navigation, sidebar, footer, comments or author box. A followed link that is only in a footer, sidebar or navigation is called out: Google lists widely distributed footer, template and widget links among its link spam examples, and an editor's link in the text counts for more.
• log: your submissions in ledger.jsonl, with a warning past five a week or when one anchor is used over and over.
• recheck: the sites in your log (submitted or live) checked again and compared with the last look: a lower verdict, a new finding, and with --to a link of yours no longer found. Kept in recheck.json. A dozen sites cost a few cents; monthly is enough.

--max-usd (competitors, cited, questions, asked) is checked against each call's estimated price before the call is made; cached answers are free. The estimates are the prices seen on the dates given, at or above what the calls cost then, not prices the providers guarantee: a call billed above its estimate (a longer answer from a model, a price raised since) can take a run past the cap by the difference. The report gives what the run cost next to the cap.

Keys

All are optional. Without them the page checks still run, and the report says what was not checked.

• GLASSER_API_KEY (glasser.ai), the first choice: spam scores, Google search history, weekly new and lost referring domains, anchors, your backlinks one by one, your competitors' referring domains, Google's AI Overviews.
• AISA_API_KEY (aisa.one), or aisa login: the same, used when there is no Glasser key; and what only AIsa sells here: the AI assistants' answers and the results for their searches (questions, asked), --deep's Ahrefs rating and Semrush numbers, and the site: search that stands in when GoAnyAPI's fails. With Glasser alone, --deep leaves Ahrefs and Semrush out and says so.
• GOANYAPI_API_KEY (goanyapi.com): rating, visits, index status, your backlink summary.

All three are referral links: signing up through them supports this project. The tool works the same with a key from each site directly.

The keys overlap on purpose. On the day of the test GoAnyAPI's search endpoints answered 502 all day; the index question went to AIsa instead and the run completed.

The rules it assumes

From submitting Agent Skills Hub to directories and lists through 2026:

1. Submit only your own project, only where it fits the list.
2. Read the contribution rules first. Some lists accept humans only.
3. One pull request per list. Never resubmit to push a queue.
4. A handful a week. A hundred directory links in a week is the pattern Google describes.
5. Vary the anchor: brand, bare URL, a plain description.
6. Never pay for a followed link. A paid placement must carry rel="sponsored" or nofollow.
7. After a listing goes live, check the link. A promise of "dofollow" is often a nofollow or a redirect.

The tool does not submit anything.

Use it from an agent: the skill and an MCP server

The skill

In Claude Code, as a plugin (it brings the MCP server below with it):

claude plugin marketplace add zhuyansen/backlink-guardclaude plugin install backlink-guard@backlink-guard

As a skill, for any agent that reads them:

npx skills add zhuyansen/backlink-guard

The skill walks an agent through the same order: find, vet, the rules, verify, audit. It runs the pinned release from PyPI (uvx backlink-guard==0.11.2), never a clone, tells you the cost and waits for your yes before any paid run, shows you where each submission goes and its exact text and waits for your yes before sending it, and treats what it reads (pages, titles, AI answers) as data, never as instructions.

MCP server

For clients that speak MCP (Claude Code, Cursor, Claude Desktop). With uv there is nothing to install first:

{  "mcpServers": {    "backlink-guard": {      "command": "uvx",      "args": ["--with", "mcp", "backlink-guard", "mcp"],      "env": {        "GLASSER_API_KEY": "...",        "GOANYAPI_API_KEY": "..."      }    }  }}

Without uv: pip install "backlink-guard[mcp]", then "command": "backlink-guard-mcp" and no args. Every key is optional, as on the command line (Keys): add AISA_API_KEY for questions, asked and --deep's Ahrefs and Semrush numbers, or in place of Glasser's. BACKLINK_GUARD_STATE puts the log and the cache where you want them. In Claude Code the plugin above brings this server with it, asks for the keys and keeps them in the system keychain.

The server gives the agent the skill's rules: the cost and a yes before each paid call, every site being considered in one vet call, and what it reads is data.

Codex, TRAE, WorkBuddy and DeepSeek Harness

The same skill and server work in these agents too. Each starts the server with uvx, so uv has to be installed.

Codex. The skill: npx skills add zhuyansen/backlink-guard -a codex (add -g for every project). The server goes in ~/.codex/config.toml, which the CLI, the IDE extension and the app share:

[mcp_servers.backlink-guard]command = "uvx"args = ["--with", "mcp", "backlink-guard", "mcp"]startup_timeout_sec = 120tool_timeout_sec = 600
[mcp_servers.backlink-guard.env]GLASSER_API_KEY = "..."GOANYAPI_API_KEY = "..."

Codex gives a server 10 seconds to start and a tool 60 seconds to answer. The first start downloads the package, and one Glasser lookup can wait in its queue for up to two minutes, so raise both as above. codex mcp add backlink-guard --env GLASSER_API_KEY=... -- uvx --with mcp backlink-guard mcp writes the same entry without the two timeouts.

TRAE. The skill: npx skills add zhuyansen/backlink-guard -a trae (TRAE CN: -a trae-cn). The server: Settings → MCP → Add → Add manually, and paste the JSON above. In TRAE CN you can instead paste this link into the browser's address bar; TRAE opens the server's configuration, where you fill in the keys you have and confirm:

trae-cn://trae.ai-ide/mcp-import?type=stdio&name=backlink-guard&config=eyJjb21tYW5kIjoidXZ4IiwiYXJncyI6WyItLXdpdGgiLCJtY3AiLCJiYWNrbGluay1ndWFyZCIsIm1jcCJdLCJlbnYiOnsiR0xBU1NFUl9BUElfS0VZIjoiIiwiQUlTQV9BUElfS0VZIjoiIiwiR09BTllBUElfQVBJX0tFWSI6IiJ9fQ%3D%3D

A key left empty counts as no key.

WorkBuddy. Put the JSON above in ~/.workbuddy/mcp.json, or paste it under Plugins → MCP servers → Configure MCP.

DeepSeek Harness. The skill: npx skills add zhuyansen/backlink-guard -a codex installs it in .agents/skills (with -g, in ~/.agents/skills), where DeepSeek Harness looks too. The server goes in ~/.dsh/cordis.patch.yml for every profile, or ~/.dsh/profiles/<name>/cordis.patch.yml for one:

- insert:    - id: backlink-guard      name: '@deepseek-ai/dsh-mcp-client'      config:        serverName: backlink-guard        transport: stdio        command: uvx        args: ['--with', 'mcp', 'backlink-guard', 'mcp']        env:          GLASSER_API_KEY: !!js process.env.GLASSER_API_KEY ?? ''          AISA_API_KEY: !!js process.env.AISA_API_KEY ?? ''          GOANYAPI_API_KEY: !!js process.env.GOANYAPI_API_KEY ?? ''        toolCallTimeoutMs: 600000

DeepSeek Harness does not pass the variables that look like keys on to a server, so the keys are named under env, read here from your environment. Keep the ?? '': without it a variable that is not set fails the entry's check and the server does not start; an empty key counts as no key. A tool gets 60 seconds unless toolCallTimeoutMs gives more. dsh web --dump-config prints the composed entry without starting anything. Tried with DeepSeek Harness 0.2.1-rc.2.

Limits

• Signals, not verdicts. A fall in visits next to a spam update is a coincidence in time until someone shows otherwise.
• A site that blocks unknown clients gives no page signals; the report says "blocked".
• Links added by script are not seen by verify.
• lost reads the strongest 1,000 backlinks at most, and a page that answers 200 for a missing page (a soft 404) looks alive to it. A suggestion matches paths, not content: read the page before you redirect to it. The same goes for a soft-lost redirect: a renamed page whose path shares little with the old one is called soft-lost until you read it.
• competitors sees what the backlink index's crawler has found: first seen is when it found a link, which can be later than the link. A linking page is looked up under a site's main domain and its www. form, so a site that links only from another subdomain shows none. A network page with a title of its own and a spam score under 50 stays in the list: vet a site before you spend time on it.
• --early counts by the day the index found each link, not the day it was placed; a link lost since counts at the age it was found. --newcomers reads 8 pages and dates 20 sites at most: its share is a sample.
• cited reads Google's US English results once: an AI Overview changes with the day, the place and the searcher.
• asked reaches the assistants through their APIs (ChatGPT through its own interface): US English, no history and no memory, not what a signed-in person sees. An answer differs from one ask to the next, so count what recurs. Gemini cites through Google's redirect, so only its sites are known, not its pages.
• --rank reads Google's and Bing's US English results once. They stand in for the indexes the assistants search, each of which has its own. On 2026-10-26 AIsa's Google results answered "Internal SE Server Error." to 21 of 67 searches: each failed search is asked once more, and one that fails twice is counted as failed, never as a search you are missing from.
• Shared accounts are read from a home page as served: an ID a tag manager or a script adds later is not seen, and a link network with no analytics at all shares nothing to read.
• Private link sellers are invisible to it.
• Visit and rating numbers are third-party estimates.
• The search history is US Google only. A site whose readers are elsewhere can fall in the US and hold at home. A fall shows once two monthly readings agree, so an update that ended last week may not show yet.
• GoAnyAPI is a young service (first released August 2026). Treat it as an enhancement.

Thanks

• Loki Yan @loki_yan_seo, whose notes on the 2026 spam updates ("follow the timeline, then go through the spam policies one by one") set the order this tool works in.
• 饭特稀 | Fentix @fantexi997 for GoAnyAPI.

MIT.