Skip to main content
Glama

AccessLens MCP

AccessLens is a local MCP server for accessibility evidence, fix verification, and consistency checks across an agent session. It scans a rendered page with axe-core, stores DOM fingerprints rather than brittle selectors alone, and compares the evidence after a fix.

Install

npx accesslens-mcp --install-browser

After it is published to npm, a client can launch it directly with npx -y accesslens-mcp. The setup command downloads Playwright Chromium once; the browser is required to scan rendered pages.

Until the npm package is published, use the GitHub package spec instead:

npx -y github:AbhishekX-dev/AccessLens-mcp --install-browser

Related MCP server: Accessibility MCP Server

Add to an MCP client

{
  "mcpServers": {
    "accesslens": {
      "command": "npx",
      "args": ["-y", "accesslens-mcp"]
    }
  }
}

Evidence Fusion Protocol

Every finding carries its raw engine evidence, DOM fingerprint, methodology, and fusion explanation. AccessLens only labels a result cross-method corroborated when separate methodologies report the same criterion on the same target. It does not convert a vote count into a false probability of truth. Verification outcomes feed a per-engine calibration record, making the fusion policy empirically auditable over time.

Human-review gate

AccessLens follows report → human review → approved remediation → verify. After every scan, an agent must call run_accessibility_review (or get_review_report after a legacy scan), show it to the human, and wait. Only explicit human decisions may be recorded with record_human_review. verify_fix refuses to run for a finding without an approved review record. This makes the intended MCP workflow human-in-the-loop; the coding agent's own system prompt should separately prohibit file edits until that approval is received.

The installed engines are axe-core, QualWeb ACT Rules (URL scans), and the AccessLens semantic analyzer. When a scan cannot preserve equivalent rendering context, that engine is explicitly marked unavailable rather than silently excluded or treated as agreement.

Tools

  • run_accessibility_review: recommended one-call workflow. Renders the target once with Playwright and returns engine status, traceable grouped issues, session checks, advisory checks, and a pending-human-review report.

  • crawl_accessibility_review: crawls same-origin anchor links breadth-first (default: 10 pages, depth 2), runs the full review on every discovered page, and returns one combined pending-human-review report. It never follows external links or invents unlinked SPA routes.

  • analyze_accessibility_evidence: scan a URL or HTML, collect evidence, and store a snapshot in a session.

  • verify_fix: re-scan an explicit updated URL/HTML and compare the original finding plus its affected subtree.

  • check_session_consistency: detect duplicate IDs, heading-level jumps, invalid landmark counts, and inconsistent accessible names.

  • get_session_report: returns outcome metrics and open findings.

  • analyze_reading_order, analyze_alt_text, analyze_control_names, and analyze_keyboard_paths: advisory modules for the roadmap checks.

Findings are returned in a versioned normalized schema designed for multiple engines and a provenance-preserving Evidence Fusion Protocol.

Security model

This is intended for local development targets. By default it accepts only http(s) URLs and blocks localhost/private-network targets unless ACCESSLENS_ALLOW_PRIVATE_TARGETS=true is set. HTML input never makes a network request.

Available Tools

12 tools
analyze_accessibility_evidenceA

Render a URL or HTML document, run axe-core plus AccessLens semantic checks (and QualWeb ACT Rules for URL scans), preserve raw per-engine provenance, and persist a pending-human-review report. Do not modify code: present get_review_report to the human first.

Inputs: url_or_html (required), session_id.

analyze_alt_textB

Advisory check for missing or generic alt text. It cannot assess semantic truthfulness without a vision model.

Inputs: url_or_html (required), session_id.

analyze_control_namesC

Advisory inventory of repeated control names and roles for human review.

Inputs: url_or_html (required), session_id.

analyze_keyboard_pathsB

Advisory static keyboard reachability check. It does not simulate full interaction flows.

Inputs: url_or_html (required), session_id.

analyze_reading_orderB

Advisory check: compare focus/tab order with rendered bounding-box order. It is not a substitute for a human reading-order assessment.

Inputs: url_or_html (required), session_id.

check_session_consistencyC

Review stored page snapshots for duplicate IDs, heading-level gaps, multiple main landmarks, and controls whose stable DOM path acquired a different accessible name across scans.

Inputs: session_id.

crawl_accessibility_reviewA

Crawl ordinary same-origin anchor links breadth-first from a start URL, within explicit page/depth limits. Runs the same Playwright, axe-core, QualWeb, structural, and advisory review on every discovered page, then returns one combined pending-human-review report. External links and unlinked SPA routes are not guessed.

Inputs: start_url (required), max_depth, max_pages, session_id.

get_review_reportC

Create the human-facing review packet: each open finding, evidence, proposed safe remediation category, and its approval status. An agent must show this report and ask for explicit human approval before editing code.

Inputs: session_id.

get_session_reportB

Return all recorded verification outcomes, the verified-fix-rate, current open findings, and scan totals.

Inputs: session_id.

record_human_reviewA

Record an explicit human decision after the human has reviewed the report. Only call this tool in response to a clear human approval or rejection; it is not permission for the agent to decide on its own.

Inputs: decision (required), finding_ids (required), note, session_id.

run_accessibility_reviewA

Preferred one-call workflow. Render a URL or HTML once with Playwright; run axe-core, AccessLens structural/session checks, QualWeb ACT Rules when an URL is supplied, and static reading-order, alt-text, control-name, and keyboard advisories. Returns one human-review report. Do not edit code; present the report and wait for explicit approval.

Inputs: url_or_html (required), session_id.

verify_fixA

Re-scan a fixed page and determine whether a previously observed finding cleared. This tool rejects verification unless that finding has a recorded human approval. Also reports new rules introduced since the original scan.

Inputs: finding_id (required), session_id, url_or_html.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.4/5.0

    Scored across 12 tools

    Disambiguation3/5

    run_accessibility_review, analyze_accessibility_evidence, and crawl_accessibility_review all trigger overlapping accessibility scans, and crawl is mainly differentiated by multi-page breadth. The descriptions help clarify intent, especially with 'Preferred one-call workflow' on run_accessibility_review, but several advisory tools also duplicate checks already included in the main review.

    Naming Consistency5/5

    Every tool uses a consistent lowercase snake_case verb_noun pattern: crawl_, run_, analyze_, get_, record_, verify_, and check_. Report, session, review, and fix nouns are used predictably, making the set highly navigable.

    Tool Count4/5

    Twelve tools is within a reasonable range for an accessibility audit server, but the count feels slightly padded by overlapping review entry points and standalone advisory checks that duplicate capabilities already present in run_accessibility_review. Consolidation would make the set tighter without losing coverage.

    Completeness4/5

    The scan-to-report-to-human-approval-to-verification lifecycle is well covered, with crawl, session consistency, and aggregate reporting adding useful breadth. Minor gaps like explicit session management or scan-scope configuration exist, but the core workflow has no dead ends.

    Maintenance

    ActivitySlowing
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers