Is MisarSEO MCP Server safe?

Findings (1)

• lowUnusually long tool description (over 2,000 characters)poison.long-description

  tool get_link_intersect: …Finds the domains that link to 2–4 of the project's competitors but NOT to the project itself. Every row is a site that has already demonstrated it will link to something in this space, which is what makes it the most actionable backlink report there is. CHARGES CREDITS — the most expensive call in the backlinks surface: one referring-domains lookup per request, 5 of them paging the project's OWN referring domains plus 1 per competitor, so one run costs up to 9 upstream requests of 1000 domains each. Every request is cached per organization+target+page, so re-running the same intersect with a different threshold, proof filter or sort is FREE — do that rather than issuing a new run, and confirm with the user before sweeping several competitor sets. UNAVAILABLE ON THE FREE DATA PROVIDER: a backlink graph needs a proprietary web-scale crawl index, so a deployment on the free provider answers with `unavailable: true`, a reason, and no rows. `totalRows`, `provenRows` and `unprovenRows` are 0 on that response and are NOT measurements — they travel beside the flag precisely so an empty table is never read as an answer. Check `unavailable` before `rows`, and never report an unavailable run as 'no opportunities exist'. `absence` has two states that must NOT be conflated. `proven` means our own referring-domain read ruled the domain out, so "does not link to us" is a measurement. `unproven` means our read was cut off at or above that domain's authority, so it may already link to us and nothing was established — never present an `unproven` row as a confirmed opportunity. `provenRows` is the headline count and excludes them by construction. `absenceBound` says which regime applies: `complete` (our whole profile was read) or `authority-floor` with the `authorityFloor` strictly above which absence is evidence. `coverage[]` is the per-target ledger — `domainsRead`, the provider's `domainsTotal`, `requestsSpent`/`requestBudget`, `truncated`, `authorityFloor`, and a per-target `unavailable`/`unavailableReason`. A competitor cell's `found: false` only means 'does not link to them' when that competitor's coverage is not `truncated`. The project's own domain is read from the project record and cannot be passed in — if the project has no domain the call fails with a validation error telling the user to set one. `minCompetitors` is clamped to the number of competitor domains that actually resolved (duplicates and the project's own domain are dropped); compare it with `minCompetitorsRequested` before calling a run empty. At most 200 rows are returned; `totalRows` is the real match count. There is deliberately no composite opportunity score and no subdomain toggle.…