CertScore.ai MCP Light

Recommended discovery entry: no-auth Streamable HTTP website privacy scans for MCP clients. Give an agent a public URL and retrieve evidence-backed observations about cookies and storage, trackers and vendors, consent controls, privacy-policy surfaces, and HTTPS/TLS signals.

CertScore.ai MCP Light is live in the GitHub MCP Registry as ai.certscore/mcp-light. It exposes four tools and requires no signup, API key, bearer token, browser login, or OAuth.

Start with MCP Light · Install in Cursor · Read the installation reference

Try it in about a minute

Add the public endpoint to Codex:

codex mcp add certscore --url https://mcp.certscore.ai/mcp/light

Then ask:

Review https://ergoveritas.com/.well-known/certscore-canary/sentinels/broad-baseline.html for website privacy risks. Explain the main findings, supporting evidence, coverage limits, and what to inspect next.

ErgoVeritas is a stable, owned canary with intentional test signals. For an ordinary review, substitute any public HTTP or HTTPS URL you are authorized to assess.

Light MCP — no authentication: start here

Light is the anonymous Streamable HTTP endpoint for first-time and low-volume agent workflows:

Light:
https://mcp.certscore.ai/mcp/light

Authentication: None
Tools: certscore_scan_site, certscore_get_scan_status, certscore_get_scan_bundle, certscore_get_report_evidence_page

Light applies session, requester-IP and shared creation limits. See the current policy at https://certscore.ai/developers/mcp#light-usage-limits. Reused eligible results do not consume quota.

Light is a free website privacy scanner and cookie checker for public websites. It can return canonical evidence and findings for pre-consent cookies and storage, trackers and vendors, cookie banners, CMP and consent controls, the jurisdiction-neutral GPC comparison, Accept and Reject Path post-action observations or explicit limitations, privacy-policy and transparency surfaces, GDPR/ePrivacy and CCPA/CPRA review signals, and HTTPS/TLS transport observations. Typical uses include release privacy preflight, public vendor-domain review, landing-page tracker inspection, audit triage, and evidence collection before human privacy review.

First-run prompt:

Review https://ergoveritas.com/.well-known/certscore-canary/sentinels/broad-baseline.html for website privacy risks. Explain the main findings, supporting evidence, coverage limits, and what to inspect next.

Canonical Light workflow:

1. Call certscore_scan_site with a public URL.
2. If a retryable error has no scanId, wait retryAfterSeconds and retry certscore_scan_site.
3. If preConsentPreview is present, it is a partial preview of checkpoint passive observations. Its counts are partial, not the full scan tally; never report them as final totals. It contains no canonical findings or score and does not replace the completed bundle.
4. If the result is queued, running, or finalizing, retain scanId.
5. Poll certscore_get_scan_status using scanId only. Never poll until scanId exists.
6. Stop polling at any terminal status. At completed or completed_limited, call certscore_get_scan_bundle before reporting full scan results, and use its final returned tally, canonical findings, and limitations; for other terminal states, follow the returned error guidance.
7. Use detail=findings for a compact finding review.
8. Use detail=evidence for finding digests, bounded external-request examples, policy-surface candidates, and references, including when no finding was projected. A request for an external script does not by itself establish tracking; use certscore_get_report_evidence_page for deeper retained report evidence.
9. If truncated, follow recommendedNextAction or increase maxBytes.
10. Summarize findings together with coverage limitations and the report URL.

Recommended bundle budgets are maxBytes=5000 for summary, maxBytes=8000 for findings, maxBytes=8000 for evidence, and maxBytes=12000 to 25000 for full. MCP Light applies a 25,000-byte response ceiling so results remain within practical client limits; larger requested budgets are reported but clamped. At the 5,000-byte floor, compact core finding rows take priority over optional inventory and duplicate envelope fields. When repeated per-finding URLs are omitted, evidenceUrlTemplate points to contentUrls.findings and the returned finding ID so the same canonical evidence endpoint remains derivable. Inspect canonicalFindingsComplete, requestedMaxBytes, effectiveMaxBytes, responseCeilingBytes, actualBytes, fullPayloadBytes, truncated, omittedSections, nextRecommendedMaxBytes, and returned report or evidence content URLs. When canonicalFindingsComplete is true, retry only if omitted envelope detail is needed.

Verification prompt:

List the available CertScore tools. Read certscore://example-report if resources are supported; otherwise call certscore_get_scan_bundle with scanId=9ba99a8c-b1ad-44c1-985f-92cef760ab40. Report the retained example’s original date and coverage. If unavailable, report that; do not create a new scan to test the connection.

Connection verification succeeds when the four Light tools are listed without an authorization page and the retained example can be retrieved with its original date and coverage. An unavailable historical example does not authorize a replacement scan. For your first website review, scan creation separately returns a stable scan ID and a new-or-reused decision; eligible reuse consumes no new-scan quota.

CertScore results are automated observations from a public-web scan. No-go, not-observed, and limited-coverage results are not proof of compliance, absence of risk, or legal status. Review the retained evidence and applicable context before relying on a finding.

Scans describe observable behavior at a point in time. Public sites may behave differently on later visits or in another execution context. CertScore findings and the CertScore score support human and agentic review; they are not legal advice, certification, or a compliance determination.

Which MCP route should I use?

The full/authenticated Streamable HTTP endpoint is https://mcp.certscore.ai/mcp. Use it after Light when the workflow requires authentication, higher volume, history, or advanced tools. Core identifiers and canonical response fields remain compatible when moving from Light to authenticated access.

Weighted read limits

MCP scan-resource reads use the same weighted, rolling policy as the direct CertScore API. Hosted MCP rejects an over-limit composite call before internal fan-out; local MCP receives the same protection from the underlying API. Poll only active status resources, stop at a terminal status, do not repeatedly retrieve terminal scan resources, and honor Retry-After before retrying. The current limits and weights are published at https://certscore.ai/developers/reference#read-rate-limits and in the x-certscore-read-rate-policy extension of the public OpenAPI documents.

Authenticated and local MCP tool reference

The sections below document the broader authenticated and local package surfaces. The Light contract exposes certscore_scan_site, certscore_get_scan_status, certscore_get_scan_bundle, and certscore_get_report_evidence_page.

• certscore_scan_site - Use for a user-requested website launch review, vendor assessment, cookie/tracker inventory, consent/GPC investigation or privacy-policy review. A public URL or domain is enough. Creates a privacy scan or reuses an eligible recent completed scan; default freshness=latest saves new-scan quota. Use refresh only for an explicitly requested fresh scan or post-fix recheck. Retain scanId, poll certscore_get_scan_status only while active at retryAfterSeconds, then read certscore_get_scan_bundle. A preliminary preConsentPreview is not final findings or totals. Coverage includes pre-consent storage, trackers, consent and CMP signals, privacy-policy disclosures, transport security and GDPR/ePrivacy or CCPA/CPRA review signals. Results are automated observations, not legal advice, certification, or a compliance determination. https://certscore.ai/developers/mcp.

  Workflow note: a new scan may include preConsentPreview when the runtime lane completes or reaches its six-second checkpoint. The preview separates captured totals from bounded returned identities. trackingVendorCount excludes infrastructure, security, and consent-management vendors, which appear in operationalVendors; the compatibility preview count is not comparable to the completed inventory's broader trackerCount. Continue with certscore_get_scan_status, never poll in parallel, and never resubmit certscore_scan_site while the scan is active. Then call certscore_get_scan_bundle at completed or completed_limited.

• certscore_get_scan - Retrieve the API v2 public-safe scan resource, including completed-limited no-go disposition, reason-specific guidance, and timing when available.

• certscore_get_scan_status - Returns lifecycle status for a stable CertScore scanId. Active responses include phase, heartbeat, estimated progress, retryAfterSeconds, and sometimes a bounded preliminary preConsentPreview. Terminal responses include completion status, CertScore score and risk metadata when available, coverage, persisted execution region and timestamps, report URL, and a next-action field. Preliminary observations are distinct from completed findings.

• certscore_get_report - Focused follow-up: retrieve a bounded Pulse report with high-signal TextContent and typed structuredContent, including customer-safe no-go messaging. For broad privacy questions, use certscore_get_scan_bundle first because it combines canonical findings, limitations, and pre-consent rows without redundant calls.

• certscore_get_evidence - Focused follow-up: retrieve a bounded public-safe evidence packet with a concise TextContent digest and typed structuredContent. For broad privacy questions, use certscore_get_scan_bundle first. Excludes raw cookie values, raw bodies, sensitive payloads, full DOM, and unredacted query values.

• certscore_get_scan_bundle - Main answer for a completed or completed-limited scan: canonical findings, next steps, supporting evidence, score, date/region, coverage and report URL. An existing scanId is enough; no new scan is created. The evidence index describes response delivery, never evidence absence. Optional next actions lead to deeper report tables, tracking workpapers and JSON/CSV exports. Use detail=evidence for request examples and policy candidates, including scans with no findings; detail=full adds bounded retained context. Preserve returned/total counts and omission metadata. Accept/Reject execution succeeded means a completed click and bounded observation; succeeded_with_confirmation additionally verifies the consent decision. Retained after-click facts remain distinct from registered decisions. Absent or failed capture remains explicitly limited. Use canonical findings for scoring effects. A later missing finding does not prove a fix. Results are automated public-web observations, not legal advice, certification, or a compliance determination.

  Count both execution success statuses; registered successful paths may omit afterAction, and missing historical execution means unavailable.

  Post-refusal observation may intentionally stop after the first qualifying non-essential activity or retained consent-signal contradiction. A confirmed observation with termination.kind=evidence_satisfied is positive evidence for the returned observation, not an inconclusive Reject Path result. coverageLimitations describe only additional behavior or persistence that was not measured.

• certscore_export_findings - Return structured findings plus completed-limited no-go disposition and guidance for downstream review or ticketing workflows.

• certscore_list_findings - Focused follow-up: list bounded API v2 public-safe findings already projected by the canonical pipeline, with matching high-signal TextContent and typed structuredContent. For broad privacy questions, use certscore_get_scan_bundle first.

• certscore_get_pre_consent_cookies_trackers - Focused follow-up: retrieve bounded row-level public-safe pre-consent cookie/tracker evidence with matching TextContent and typed structuredContent. For a new broad request such as checking a site for pre-consent tracking, use certscore_scan_site then certscore_get_scan_bundle first.

• certscore_explain_finding - Explain one projected finding with public evidence, caveats, reviewer next steps, and reason-specific no-go context when applicable.

• certscore_get_latest_domain_scan - Retrieve the latest eligible API v2 public-safe scan for a domain.

Status provenance fields remain distinct: provenance.retrievalMode describes the current read, provenance.creationDecision reports the retained original new/reused decision or unknown, and provenance.scanAgeSeconds reports numeric age when available. A scan-ID lookup alone never proves reuse.

At tight bundle budgets, short canonical nextStep actions remain only when they fit without displacing a finding. Longer actions and evidence detail remain available through the returned canonical URLs or a larger complete tier.

• certscore_get_latest_domain_pre_consent_cookies_trackers - Focused follow-up: retrieve bounded row-level public-safe pre-consent cookie/tracker evidence from the latest eligible scan for a domain, with matching TextContent and typed structuredContent. For a broad current-site review, use certscore_scan_site then certscore_get_scan_bundle first.

The initial MCP surface intentionally does not include account scan browsing or scan comparison tools.

Scan Timing Fields

MCP tools backed by API v2 scan resources return scan timing when CertScore has enough timing evidence:

• startedAt
• completedAt
• scanTimeSeconds

This applies to certscore_scan_site when it returns an API v2 scan resource or job, certscore_get_scan, and certscore_get_scan_status when called with a scanId. scanTimeSeconds: null means timing is unavailable or incomplete and should not be displayed as 0.

Completed Light results are canonical. certscore_scan_site, terminal certscore_get_scan_status, and certscore_get_scan_bundle return the same score, risk level, coverage, and timing fields. Scores include scoreStatus, scoreVersion, and scoreUpdatedAt; a scan remains finalizing until the persisted canonical report projection is ready, so a completed response always carries scoreStatus: "final".

No-go results lead with the retained access blocker, “Not scored,” evidence excerpt, next action, and retry guidance in MCP text. Structured responses retain resultDisposition, noGo, null score/risk, and empty findings. Tight bundle budgets preserve that blocker and remedy before optional lane diagnostics; omitted diagnostics are listed in mcpMetadata.omittedSections. A no-go full request does not fabricate a full report.

The value is labeled CertScore score, never a compliance score. It covers observable public-web scan signals only. Clients must not infer technologies absent from the returned evidence, compare the value with a hypothetical compliant baseline, or infer legal compliance status.

Every failed, expired, or rate_limited status includes a bounded error object with code, message, retryable, retryAfterSeconds, and recommendedNextAction.

Light Review Follow-ups

Completed bundles include an evidence index and optional next actions for task-relevant report evidence and tracking JSON/CSV exports. Delivery labels describe the response, not evidence absence. Summarize whether the result was new or reused. Retain the baseline scan ID, date and execution region for a later explicitly requested fresh recheck; missing findings do not verify a fix. Offer remediation or the same review for another user-selected site only when useful. Light provides the optional certscore_launch_review, certscore_compare_scans and certscore_remediation_checklist prompts and certscore://project-instructions resource. certscore://example-report tests retrieval without creating a scan.

Light Bundle Detail and Byte Budgets

Every bundle declares its selected detail mode. summary returns the overview, canonical result, up to five compact projected finding objects, compact row-level pre-consent cookie/tracker evidence, coverage, limitations, counts, and report URL. This lets conversational MCP clients enumerate projected consent-control, CMP-context, transport, GDPR/ePrivacy transparency, social/media embed, accessibility, disclosure, and other returned review signals without needing a second tool; only categories actually present in canonical projections are returned. Findings and inventory sections each declare total, returned, and truncated. Each compact finding retains its public API v2 evidence anchor. Each inventory row includes cookie/tracker identity, cookie names where present, vendor, purpose, category, first-observed time, domains, evidence classification, and confidence. findings raises the default finding allowance to 20. evidence adds bounded evidence digests and references. full adds every available non-duplicated section subject to the effective byte budget; findings, top findings, and transport security are returned once in their canonical top-level bundle sections. When fullPayloadBytes exceeds the Light response ceiling, use the canonical content URLs instead of retrying above the ceiling.

TextContent is capped at 8,000 characters and uses short plain-text lines instead of large tables or nested JSON. It presents canonical overview facts and projected findings before the row inventory so one evidence family cannot crowd out the rest. The full profile's structured bundle defaults to 50,000 bytes and accepts a caller-selected 5,000-200,000-byte bound. Light defaults to and applies a 25,000-byte response ceiling. If either representation must omit returned items, it states that explicitly and preserves totals and truncation metadata.

mcpMetadata always includes requestedMaxBytes, effectiveMaxBytes, responseCeilingBytes, responseBudgetClamped, actualBytes, fullPayloadBytes, truncated, canonicalFindingsComplete, truncationReason, omittedSections, deduplicatedSections, nextRecommendedMaxBytes, omittedContentAvailableViaUrl, and contentUrls. Under byte pressure, optional detail and inventory rows are reduced before compact core finding rows. canonicalFindingsComplete separates complete projected findings from a partially omitted envelope. nextRecommendedMaxBytes is the rounded size needed for the complete requested tier rather than an incremental retry step; when the complete tier exceeds the MCP ceiling, the response directs the agent to an available report/evidence URL.

Input-validation errors remain MCP -32602 errors and also include structured invalid_arguments details with the affected field and a safe next action. Error results use concise text plus machine-readable details. Successful results put the typed result in structuredContent; scan bundles also render a bounded row-level evidence summary in TextContent for client compatibility.

The Light workflow is:

1. Call certscore_scan_site with a public URL.
2. If a retryable error has no scanId, wait retryAfterSeconds and retry certscore_scan_site; do not poll.
3. If the result is queued, running, or finalizing, retain scanId and poll certscore_get_scan_status with only that ID.
4. Stop polling at any terminal status. For completed or completed_limited, call certscore_get_scan_bundle.
5. If the bundle is truncated, follow recommendedNextAction, increase maxBytes, or open a listed content URL.
6. Summarize the CertScore score, risk, pre-consent rows, findings, coverage, limitations, and report URL without treating no-go, not-observed, or limited coverage as proof of compliance.

Completed-Limited No-Go Results

No-go scans are usable terminal results, not transport failures. Relevant tools retain status: "completed_limited", resultDisposition: "no_go", the stable reason code, customer-safe title and explanation, limitationKind attribution, retry guidance, and a bounded evidenceExcerpt when retained. Unknown future reasons use generic customer copy while remaining structured as reasonCode: "unknown".

Hosted MCP — OAuth

OAuth-capable MCP clients can connect to:

https://mcp.certscore.ai/mcp

Discovery endpoints:

https://mcp.certscore.ai/.well-known/oauth-protected-resource/mcp
https://certscore.ai/.well-known/oauth-authorization-server

The full hosted service uses OAuth authorization code with PKCE. Members of active workspaces can connect through registered OAuth clients across plans with scan:read, scan:create and mcp. Existing limits apply. Canonical eligibility, authorization and reconnect behavior: https://certscore.ai/developers/mcp#hosted-oauth-start. Local scoped API-key policy is separate.

For low-volume agent discovery without account or OAuth setup, use the unauthenticated endpoint:

https://mcp.certscore.ai/mcp/anonymous

For the simplest no-account workflow, use the Light quickstart at the top of this document. OAuth metadata applies only to https://mcp.certscore.ai/mcp.

Local MCP — scoped API key configuration

Install with Homebrew on macOS:

brew tap ergoveritas1-alt/certscore https://github.com/ergoveritas1-alt/certscore.ai
brew install --cask certscore-mcp

The cask installs the prebuilt MCP command for users who prefer a persistent local binary.

Use the installed command from an MCP client:

{
  "mcpServers": {
    "certscore": {
      "command": "certscore-mcp",
      "env": {
        "CERTSCORE_API_KEY": "YOUR_TOKEN",
        "CERTSCORE_BASE_URL": "https://certscore.ai"
      }
    }
  }
}

Run from this monorepo for local development:

CERTSCORE_API_KEY=... pnpm mcp:certscore

Generate a scoped preview key after applying DB migrations:

pnpm db:migrate
pnpm mcp:certscore:generate-key -- --name "CertScore MCP preview"

Run the built package directly after local build:

CERTSCORE_API_KEY=... certscore-mcp

Optional:

CERTSCORE_BASE_URL=https://certscore.ai
CERTSCORE_REQUEST_TIMEOUT_MS=300000

CERTSCORE_API_KEY should be a scoped CertScore API token for the workspace or preview user. The MCP server passes it to Pulse as a bearer token and does not persist it.

Local MCP — scoped API key access

Stdio API keys use pulse:read and mcp; creating scans additionally requires pulse:scan. Hosted OAuth uses scan:read scan:create mcp under the registered-client active-workspace policy. See https://certscore.ai/developers/mcp#hosted-oauth-start for current eligibility and verified client availability.

Verify Install

certscore-mcp --version
certscore-mcp --help
CERTSCORE_API_KEY=... certscore-mcp doctor
CERTSCORE_API_KEY=... certscore-mcp doctor --check-auth

The doctor command checks binary startup, version output, Node.js runtime compatibility, the configured CertScore base URL, API v2 health, and API key presence. Add --check-auth to validate the credential against /api/v2/auth/check without creating a scan. It does not print secrets or inspect raw scanner artifacts.

No-account agent scan path

Agents that cannot create an account or configure OAuth can use the public API v2 scan path without an Authorization header:

curl -X POST https://certscore.ai/api/v2/scans \
  -H "Content-Type: application/json" \
  -d '{"url":"https://ergoveritas.com/.well-known/certscore-canary/sentinels/broad-baseline.html","freshness":"latest","scanFrom":"eu_ie"}'

New anonymous scans are limited to 20 per requester IP per UTC day. Reusing an eligible recent result does not consume the quota. Poll the returned status resource, then retrieve findings or evidence. Contact support@certscore.ai for a higher-volume allowance, including when reuse serves the current request.

MCP Client Examples

Claude Desktop-style config:

{
  "mcpServers": {
    "certscore": {
      "command": "certscore-mcp",
      "env": {
        "CERTSCORE_API_KEY": "YOUR_TOKEN",
        "CERTSCORE_BASE_URL": "https://certscore.ai"
      }
    }
  }
}

Cursor config:

{
  "mcpServers": {
    "certscore": {
      "command": "certscore-mcp",
      "env": {
        "CERTSCORE_API_KEY": "YOUR_TOKEN",
        "CERTSCORE_BASE_URL": "https://certscore.ai"
      }
    }
  }
}

Windsurf or generic stdio MCP client config:

{
  "mcpServers": {
    "certscore": {
      "command": "certscore-mcp",
      "env": {
        "CERTSCORE_API_KEY": "YOUR_TOKEN",
        "CERTSCORE_BASE_URL": "https://certscore.ai"
      }
    }
  }
}

Local repo config for contributors:

{
  "mcpServers": {
    "certscore": {
      "command": "pnpm",
      "args": ["mcp:certscore"],
      "cwd": "/path/to/WC01",
      "env": {
        "CERTSCORE_API_KEY": "YOUR_TOKEN",
        "CERTSCORE_BASE_URL": "https://certscore.ai"
      }
    }
  }
}

Agent Workflow

1. Call certscore_scan_site with a public URL. A reused eligible result may be completed immediately; a new scan returns its stable scanId and may return a partial preview when the runtime lane completes or reaches its six-second checkpoint.
2. If preConsentPreview is present, distinguish captured totals from bounded returned identities. Use trackingVendorCount for non-operational tracking vendors and keep operationalVendors separate; do not compare the compatibility preview trackerCount with the completed inventory's broader trackerCount. It is not a finding, score, or terminal result.
3. If status is queued, running, or finalizing, retain the returned scanId and poll certscore_get_scan_status using only that ID. Stop polling at completed, completed_limited, failed, expired, or rate_limited.
4. For completed or completed_limited, call certscore_get_scan_bundle with its default detail: "summary" before reporting full scan results. Use its final returned tally, canonical findings, and limitations. Use findings, evidence, or full only when progressively deeper bounded context is required.
5. Summarize the canonical score, risk, findings, coverage, limitations, report URL, and next action. Never treat the preview, no-go, not-observed, or limited coverage as proof of compliance.

Canonical first-run prompt:

Scan these public URLs. For each one, call certscore_scan_site. If preConsentPreview is present, treat it as a partial preview whose counts are checkpoint-only and partial, not the full scan tally; never report them as final totals, and continue the workflow. If status is queued, running, or finalizing, retain scanId and poll certscore_get_scan_status using only scanId. Stop polling at completed, completed_limited, failed, expired, or rate_limited. For completed or completed_limited scans, call certscore_get_scan_bundle with detail=findings before reporting full scan results. Use the bundle for the final returned tally, canonical findings, and limitations. If truncated, follow recommendedNextAction or increase maxBytes. Report the canonical score, risk level, coverage, findings, limitations, report URL, and next action. Never treat preConsentPreview, no-go, not-observed, or limited coverage as proof of compliance.

4. Call certscore_get_report, certscore_get_evidence, certscore_list_findings, or certscore_get_pre_consent_cookies_trackers only when the task needs a dedicated view.
5. Call certscore_explain_finding when a reviewer needs evidence and caveats for a specific finding.
6. Call certscore_get_latest_domain_scan or certscore_get_latest_domain_pre_consent_cookies_trackers when the user asks for latest eligible public data for a domain.

certscore_scan_site reports whether the result was reused, why the freshness decision was made, whether anonymous quota was consumed, the remaining daily allowance, the UTC reset time, and the recommended next tool.

With freshness: "latest", CertScore reuses an eligible scan completed within the last 24 hours for the same normalized target and scan region. A reusable result must have completed usable page coverage and must not be an early-loss, no-page, or otherwise non-reusable limited result. Reuse does not consume anonymous quota. freshnessDecision states whether a recent result was reused or a new scan was queued; reusedScanAgeSeconds reports the reused result's age.

Provenance keeps the current retrieval separate from the original creation decision. retrievalMode=creation_response means the result came from certscore_scan_site; retrievalMode=scan_id_lookup means a later tool fetched the retained scan by ID. creationDecision is new_scan or reused_scan only when the response retains that fact, and otherwise is unknown. Do not translate scan_id_lookup into a reused-scan claim. scanAgeSeconds reports a nonnegative age when a retained age or completion timestamp is available.

{
  "tool": "certscore_get_pre_consent_cookies_trackers",
  "arguments": {
    "scanId": "00000000-0000-4000-8000-000000000123"
  }
}

{
  "tool": "certscore_get_latest_domain_pre_consent_cookies_trackers",
  "arguments": {
    "domain": "ergoveritas.com",
    "scanFrom": "eu_ie"
  }
}

When summarizing table data, group rows by vendor, purpose, and host unless the user asks for row-level JSON. Treat MCP outputs as automated public-web observations for human and agentic review. They are not legal advice, certification, or a compliance determination. MCP tools must not infer findings from raw labels, raw network events, missing data, or display-only context.

Live Smoke

CERTSCORE_API_KEY=... pnpm mcp:certscore:smoke

Optional:

CERTSCORE_MCP_SMOKE_URL=https://ergoveritas.com/.well-known/certscore-canary/sentinels/broad-baseline.html

Without CERTSCORE_API_KEY, the smoke script exits successfully with a skip message.

For the full production operator smoke, run from the WC01 repo:

CERTSCORE_ALLOW_PAID_ECS_SMOKE=1 pnpm ops:smoke:mcp-cli-production

This verifies the Homebrew-installed certscore-mcp command against live https://certscore.ai. It first requires the installed CLI version to match the workspace version, then creates a short-lived preview key, stores only the hash in production through the approved ECS/Fargate path, checks required tools, requires non-empty findings and pre-consent cookies/trackers rows, and revokes the temporary key afterward. The explicit cost opt-in is required because this separate CLI check starts one-off Fargate tasks. Use pnpm ops:smoke:mcp-production for the hosted, retained-scan, read-only canary.

Troubleshooting

• Unexpected OAuth in Codex: remove the server and add it again with the exact Light URL https://mcp.certscore.ai/mcp/light. Do not configure a bearer token.
• Successful Light connection: Streamable HTTP initialization completes, no authorization page opens, and tools/list returns exactly certscore_scan_site, certscore_get_scan_status, and certscore_get_scan_bundle.
• Missing scan ID: retry certscore_scan_site only when the error is retryable. Never call certscore_get_scan_status until scanId exists.
• Rate limited: follow retryAfterSeconds and recommendedNextAction, wait for the returned UTC reset, or reuse an eligible result.
• Reused result: report that the eligible prior scan was reused and quota was not consumed.
• Truncated bundle: first inspect canonicalFindingsComplete; when true, retry only for omitted envelope detail. Otherwise follow nextRecommendedMaxBytes, raise maxBytes, or open a returned report or evidence content URL.
• Invalid URL: correct the field named by the structured invalid_arguments response and retry with a public HTTP or HTTPS URL.
• Limited result versus failure: completed_limited, no-go, not-observed, and limited coverage are observations only, never proof of compliance. failed, expired, and connection errors are failures with retry guidance.
• Command not found: run the Homebrew install again and confirm Homebrew's bin directory is on PATH.
• Missing API key: set CERTSCORE_API_KEY in the MCP client environment and rerun certscore-mcp doctor.
• Bad token: rotate the key or request a scoped API/MCP key from support@certscore.ai.
• API unreachable: check CERTSCORE_BASE_URL and verify https://certscore.ai/api/v2/health.
• Homebrew tap stale: run brew update and reinstall certscore-mcp.
• Old cached release: run brew reinstall --cask certscore-mcp after updating the tap.

Runbook

See docs/certscore-mcp-homebrew-release.md for Homebrew release steps and docs/certscore-mcp-preview-runbook.md for key issuance, smoke testing, deploy verification, and scan-to-report guardrails.

Hosted request diagnostics

Hosted MCP retains bounded, redacted previews of explicitly submitted tool arguments and client metadata for operational diagnostics. The admin “What the caller sent” popup distinguishes current arguments, per-request metadata, and initialization metadata. It records reasons for sensitive-field, sensitive-content, depth, text, field-count, and byte-budget omissions. These caller-declared values are not verified claims or a copy of the surrounding conversation.

The entire request-details record remains capped at 4 KB with the existing 90-day event retention. Previews contain at most 24 fields and 300 characters per string. Authentication headers, cookies, known credential fields, and chat-history/message collections are excluded; URL previews retain only the origin. Task-context question summaries retain their explicit sharing/source requirements and cannot be recovered through the generic preview path. The generic preview may show short text explicitly submitted in separate arguments such as reason, notes, or prompt; it does not fetch the user's chat.

Question context from a prior request is shown separately, with a source link, only when the retained caller, session, scan, provider and entrypoint match and the source precedes the current call by at most 24 hours. Missing identities, filtered traffic and invalid context fail closed. No inherited context is written back to current events, and no scanning, findings or scoring behavior changes.

Estimated additional cost at the September 2026 observed request volume: less than $0.10/month, using existing storage and record limits. No model calls, infrastructure capacity changes, or longer retention are introduced.

Complete report evidence JSON (OAuth and Light)

Use certscore_get_scan_bundle for a concise summary. For all fields of the displayed single-page or full-site report, use certscore_get_report_evidence_page({scanId}). Continue with {scanId, cursor: pagination.nextCursor} until pagination.complete is true. This adds a fourth Light tool; the three-tool scan → status → bundle workflow remains the default.

Each page contains entries with RFC 6901 JSON Pointer path and JSON value. Apply them in order, creating parent containers before children. An empty path replaces the root. For an oversized string, concatenate value by zero-based stringPart through stringParts before assigning it. Treat paths as data; use own properties when reconstructing objects to prevent prototype pollution. All pages must have the same snapshot; HTTP 409 means restart without a cursor and discard the old partial export.

The export preserves canonical findings, evidence tables, policy excerpts, inventory, consent/action/GPC evidence and coverage as present in the report projection. A complete export does not mean complete scan observation: retained samples, unavailable evidence and report limitations remain authoritative. Raw scanner artifacts outside the report and inline image binary bytes are excluded. Full-site data appears under fullSiteReport, including all page/resource rows, services, forms and retained fields. Available form snapshots have download URLs under fullSiteReport.collectionSurfaces.rows[].snapshot.url; use the OAuth bearer credential for workspace images, or no credential for eligible anonymous public scans. These downloads retain the existing provenance and image-safety checks. Withheld/unavailable snapshots are not promoted. OAuth can read its workspace’s reports and eligible public scans; Light can read only eligible anonymous public scans. No scan is created by this tool. Existing read throttles apply; honor Retry-After.

Direct JSON endpoint: GET /api/v2/scans/{scanId}/report-evidence?cursor={nextCursor}. TypeScript SDK: client.getReportEvidencePage(scanId, {cursor}).

• certscore_get_connection_status - Read current authenticated connection mode, granted scopes, workspace access, rolling scan quota and recovery action. No scan ID is needed and no scan is created. Use this to diagnose read-only access or quota limits; reconnect only for expired, revoked or expanded access.

• certscore_get_report_evidence_page - For a focused question, use section=consent, gpc, policy, tracking, transport or forms. This returns selected retained report sections plus shared scan, score, findings and coverage context. Preserve section on cursor continuation; unselected or not-returned fields are not evidence of absence. Use workpaper=tracking separately for the starting-page tracking inventory, privacy choices/notices and GPC evidence, with JSON and CSV downloads. Never combine section and workpaper. The workpaper selector also applies to every continuation request. Otherwise retrieve scan report display content as paginated JSON, without internal diagnostic JSON downloads. The response also offers a single-file full JSON download; private JSON download links expire after five minutes and need no OAuth header; use pagination if your host blocks file downloads. Repeated display records use reportContentRef JSON Pointers. Includes evidence tables, full-site page and resource inventories, all retained additional-page form fields, form snapshot download references, and retained limitations. Snapshot images are downloaded separately from the returned URLs, with OAuth bearer authentication for workspace scans. Available on OAuth and Light. Start with scanId; follow pagination.nextCursor until complete. Pages share a snapshot; restart if it changes. Each entry has a JSON Pointer path and value; oversized strings use numbered parts. Export completion is not complete observation coverage. Use the concise scan bundle for summaries; use this tool for exhaustive report evidence. No new scan is created.

Report-evidence export pages contain up to 64 KB of JSON entries and cost one terminal-read unit on both hosted MCP and the API. Existing 120-unit/10-minute and daily limits remain enforced, including repeated cursors. Full evidence reports, findings exports and bundles remain four units. If a larger export reaches a limit, retain the last nextCursor, wait for Retry-After, and resume rather than restarting.

Full report JSON download

certscore_get_scan_bundle remains the default for concise summaries. For a full report export, call certscore_get_report_evidence_page once: download.url returns the entire report display JSON in one HTTP response, with its byte size. Workspace download URLs contain a report-only capability that expires after five minutes. Fetch the returned URL directly without OAuth headers. Treat private links as confidential; do not publish them. Request a fresh page for a new link after expiry. Eligible public reports remain anonymously downloadable. If the host cannot fetch authenticated files, follow pagination.nextCursor through MCP instead. This does not require adding another connector.

Exports exclude diagnostic JSON downloads and internal runtime graphs. Repeated records use reportContentRef RFC 6901 pointers within the exported document; resolve these after loading it. Form fields, inventory rows, coverage limitations, and available snapshot links remain included. Image bytes are separate downloads. The download uses the existing report read quota and creates no new scan or stored export artifact. Actual host download support must be tested separately; a returned URL is not proof that Cursor or Claude can retrieve it.

Privacy and tracking workpapers (0.2.24)

certscore_get_scan_bundle includes privacyAuditSummary when verified retained evidence is available: observed privacy-choice controls, notice topics, source hash and retained counts. detail=full also includes privacyAuditEvidence. Byte-budget omissions are named in mcpMetadata.omittedSections; absence of a section is not evidence that a control or notice was absent.

Use the existing reader for the starting-page inventory and workpaper:

{"scanId":"00000000-0000-4000-8000-000000000123","workpaper":"tracking"}

Pass those arguments to certscore_get_report_evidence_page. It returns paginated JSON entries, download.url for the JSON workpaper and download.csvUrl for the inventory CSV. Preserve workpaper on every cursor continuation. This works on OAuth and Light with their existing workspace/public access boundaries and read quotas; no scan is created. Keep five-minute private download links confidential.

Lead GPC summaries with returned observation facts and gpcResponse.activityComparison request counts when present. Keep the matched duration with those counts. Report Do Not Sell/Share control presence and retained notice passages as observations; do not infer opt-out effectiveness, notice adequacy, sale/sharing, or GPC honoring. Historical records remain unchanged and the score is the existing CertScore score.

Focused report sections (0.2.26)

For a specific question, call certscore_get_report_evidence_page with the existing scanId and one section: consent, gpc, policy, tracking, transport, or forms. Preserve that selector with every returned cursor. Shared scan identity, score, findings and coverage remain unchanged. Selection happens before deduplication, so references stay inside the selected document. selection.notReturnedFields describes missing delivery, never verified absence. JSON download links preserve the selected section. Use workpaper="tracking" separately for the starting-page workpaper and CSV; never combine section and workpaper. No additional scan, model call or evidence capture is performed.