pseolint

@pseolint/mcp

MCP server for pseolint: audit pSEO sites by template from AI coding assistants.

An MCP (Model Context Protocol) server that exposes pseolint v0.7.5 auditing tools to AI coding assistants like Claude Code, Claude Desktop, Cursor, and Windsurf.

All tools are namespaced with a pseolint_ prefix (pseolint_audit_site, pseolint_explain_score, pseolint_check_page_technical, pseolint_orchestrate_audit) so they don't collide with other MCP servers, and each returns both human-readable text and machine-readable structuredContent.

What's new in v0.6: per-template output in tool responses

audit_site now returns a templates array alongside the existing findings list. AI clients get the per-template view automatically; no parameter changes needed.

{

"verdict": "concerning", "risk": 60, "templates": [ { "signature": "/listing/:slug", "totalUrls": 8201, "auditedUrls": ["https://example.com/listing/foo", "..."], "verdict": "concerning", "risk": 60, "variance": { "uniformityScore": 0.85, "topDriver": { "ruleId": "spam/thin-content", "fireRate": 0.8 } } }, { "signature": "/category/:slug", "verdict": "ready", "risk": 12 } ], "findings": [...] }

When iterating templates, the topDriver rule + fireRate is the most actionable signal: "8/10 samples fail spam/thin-content" tells the LLM which template is broken and what to fix. The findings flat list remains available for per-URL drill-down.

Design rationale: docs/superpowers/specs/2026-05-04-pseolint-v0.6-audit-as-template-reframe.md

What's new in v0.5.2: credibility layer

• 4 new content-quality rules in the underlying engine (consumed by audit_site and orchestrate_audit): content/title-uniqueness, content/heading-structure, content/image-alt-text, tech/og-completeness. Findings appear in tool output as standard rule findings.
• `audit_site.authorityScore` parameter (0-100): bring-your-own-DA. >= 80 shifts the verdict one tier lenient on established brands; <= 30 shifts one tier stricter on newer/lower-authority operators. Raw risk number unchanged.
• `audit_site.sampleSeed` parameter: deterministic sampler. Same seed = same audit = same verdict, run after run. AI assistants asking the user to confirm a finding can re-audit reproducibly.
• `spam/doorway-pattern` cluster collapse: a 276-pair finding on a catalog directory now arrives as one cluster line per group, not 276 line items eating the LLM's context window.
• Findings stay actionable: info-severity findings are capped per category bucket so they can't accumulate to tank a verdict on their own; the LLM gets the actual signal, not noise.
• Calibrated against reputable in-production pSEO sites; trade-offs and limitations documented at pseolint.dev/methodology.

What's new in v0.5

`orchestrate_audit` tool. Drives an LLM through 25 deterministic audit tools and produces a fix manifest with concrete copy-paste patches (rewritten H1s, JSON-LD blocks, robots.txt diffs). Use when a user wants concrete fixes, not just a list of issues.

Conservative MCP defaults: $2 / 60 tool calls / 180 seconds wall (vs CLI's $5 / 100 / 300). Two output modes: summary (terse text for chat UI) and json (full manifest + validation + diff). Each invocation reports actual USD spend. Patches that fail deterministic validators are dropped from the manifest and surfaced separately.

Example prompt: "Use the pseolint_orchestrate_audit tool to run an AI-native audit of https://example.com with concrete fix proposals."

Safety defaults (v0.3.3+)

All three tools default to safeMode: "saas"; AI assistants running in end-user environments can't be tricked into scanning AWS/GCP metadata endpoints, localhost, or RFC1918 networks via a malicious URL argument. Specifically:

• guardSsrf: true: DNS-validated private-range check on the source URL, sitemap entries, redirect hops, and discovered links
• respectRobotsTxt: true: sitemap URLs Disallow'd by the target's robots.txt are skipped instead of crawled
• Tighter maxFetchBytes (10 MB) and maxCrawlDiscovered (2000) caps

Tools

pseolint_orchestrate_audit (v0.5)

Drive an LLM through 25 audit tools and produce a fix manifest with concrete patches. Use when a user wants paste-able fixes (not just a list of issues). Costs ~$1-3 per audit on managed Anthropic.

Parameters:

• domain (required): URL of the site to audit (e.g. https://example.com)
• maxCostUsd: Hard USD cap (default 2)
• maxToolCalls: Hard tool-call cap (default 60)
• maxWallSeconds: Hard wall-clock cap (default 180)
• format: summary (terse text) or json (full manifest + validation + diff)

Returns: text summary with verdict + categories + top-3 patches per bucket (or full JSON when format: "json"). Validation failures listed separately so the LLM-host conversation stays grounded in what actually shipped.

Example prompt: "Use pseolint_orchestrate_audit on https://example.com with format=summary"

pseolint_audit_site

Run a full pseolint audit on a URL or directory path. Returns the site-level verdict + risk score, a templates array (per-template verdicts + variance metrics), and all per-URL findings with actionable fix suggestions.

Parameters:

• source (required): URL or directory path to audit
• threshold: Score threshold for pass/fail (default: 40)
• sampleSize: Audit a random subset of N pages (0 = all)
• format: Output format: console or json (default: console)
• authorityScore: 0-100 domain authority hint. ≥80 shifts verdict one tier lenient; ≤30 shifts one tier stricter. Raw risk number unchanged.
• contentEffort (boolean, default false): enable the AI content-effort signal: an LLM judges a 0-100 originality/effort score from sampled page text that moderates the verdict ±1 tier. Requires ANTHROPIC_API_KEY in the MCP server's environment; no-ops safely without one. Adds a few cents of LLM cost per audit.
• sampleSeed: Integer seed for deterministic stratified sampling. Same seed = same audit = same verdict.

Returns (v0.6 shape): verdict, risk, categories, templates: Template[], findings: RuleResult[]. AI clients should iterate templates first; topDriver.ruleId + fireRate is the most actionable per-template signal. Use findings for per-URL drill-down.

Example prompt: "Audit my site at http://localhost:3000; show me the per-template breakdown"

pseolint_explain_score

Run an audit and get a human-readable explanation of what's driving the SpamBrain Risk Score, including category breakdowns, top issues, and prioritized fix suggestions.

Parameters:

• source (required): URL or directory path to audit
• threshold: Risk threshold for the pass/fail verdict (default: 40)
• authorityScore: 0-100 domain authority hint. ≥80 shifts verdict one tier lenient; ≤30 shifts one tier stricter. Raw risk number unchanged.
• contentEffort (boolean, default false): enable the AI content-effort signal: an LLM judges a 0-100 originality/effort score from sampled page text that moderates the verdict ±1 tier. Requires ANTHROPIC_API_KEY in the MCP server's environment; no-ops safely without one. Adds a few cents of LLM cost per audit.
• sampleSeed: Integer seed for deterministic stratified sampling.

Example prompt: "Explain why my site's SpamBrain score is high"

pseolint_check_page_technical

Check a single page URL for per-page technical SEO issues (canonical, Open Graph, JSON-LD, robots, meta, thin content, author signals). Does not run cross-page rules; use pseolint_audit_site for those.

Parameters:

• url (required): Full URL of the page to check

Example prompt: "Check https://yoursite.com/templates/california-llc for technical SEO issues"

Resources

The server also exposes pseolint's rule knowledge as read-only MCP resources, so an assistant can explain a finding without guessing or fetching the web. The resources are keyed by ruleId, the same identifier that appears on every audit finding (e.g. spam/thin-content), so they line up 1:1 with tool output.

• `pseolint://rules`: a JSON index of every documented rule (ruleId, title, and the per-rule resource URI). Read this first to discover what's available.
• `pseolint://rules/ `: one Markdown resource per rule, e.g. pseolint://rules/spam/thin-content. Each contains the rule's link, a What it detects section, and a How to fix checklist.

Typical flow: run pseolint_audit_site, take a ruleId off a finding, then read pseolint://rules/ to ground the explanation and fix advice in pseolint's own rule documentation.

Remote server (hosted, zero-install)

Prefer not to install anything? Point your MCP client at the hosted endpoint:

https://pseolint.dev/mcp

It serves the three read-only audit tools (pseolint_audit_site, pseolint_explain_score, pseolint_check_page_technical) with no signup; anonymous use is rate-limited. Create an API key in your pseolint.dev dashboard and send it as Authorization: Bearer to raise your limits. The AI-orchestrated pseolint_orchestrate_audit tool is available only via the stdio package (below) or the CLI for now.

Example client config:

{ "url": "https://pseolint.dev/mcp", "headers": { "Authorization": "Bearer " } }

Installation

Claude Code

claude mcp add pseolint -- npx @pseolint/mcp

Claude Desktop

Add to claude_desktop_config.json:

{

"mcpServers": { "pseolint": { "command": "npx", "args": ["@pseolint/mcp"] } } }

Cursor

Add to .cursor/mcp.json:

{

"mcpServers": { "pseolint": { "command": "npx", "args": ["@pseolint/mcp"] } } }

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

{

"mcpServers": { "pseolint": { "command": "npx", "args": ["@pseolint/mcp"] } } }

VS Code (GitHub Copilot)

Add to .vscode/mcp.json:

{

"servers": { "pseolint": { "command": "npx", "args": ["@pseolint/mcp"] } } }

What It Checks

65 rules, grouped by theme (rolled up into 4 scored categories: integrity, discoverability, citation, data):

• SpamBrain Risk: near-duplicate detection, entity-swap doorway pages, thin content, boilerplate ratio
• Content Quality: unique value per page, heading/meta uniqueness, E-E-A-T signals
• Internal Linking: orphan pages, dead ends, cluster connectivity, link depth
• Technical SEO: canonical consistency, sitemap completeness, robots.txt conflicts
• Structured Data: JSON-LD validation, required fields, schema consistency
• Cannibalization: URL pattern conflicts (title-overlap and keyword-collision were dropped in v0.4 for high false-positive rates)

Links

• GitHub
• npm: pseolint
• npm: @pseolint/core

License

MIT