superagent_

Context Guardrails

Context Guardrails score web pages, email, files, agent skills, GitHub hosted MCP repositories, and registry packages before agents trust them.

Agents fetch packages, load web pages, install MCP servers, and clone repositories without reviewing any of it. Context Guardrails score every piece of external context before the agent trusts it, so agents keep full capability without inheriting untrusted input.

This feature complements Runtime Guardrails, which decide what an agent may do. Context Guardrails decide what an agent may consume.

What gets scored

Entity Status What Superagent checks
Web pages Available Domain identity, page behavior, and page content before the agent reads the page
Email Available Sender identity, phishing and injection patterns, attachments, and outbound links
Files Available Text and PDF contents, source identity, redirects, type consistency, encoded instructions, and prompt injection
Agent skills Available GitHub publisher identity, requested capabilities, scripts, hidden instructions, exfiltration, and prompt injection
MCP repositories Available Identity, tool descriptions, schemas, instructions, hooks, secrets, and executable behavior in a public GitHub repository
Packages Available Registry artifacts for npm, PyPI, Go, RubyGems, and GitHub Actions, using the same supply chain pipeline as pull request scans
Repositories Covered by repository red teaming Agent-driven source review for vulnerabilities, backdoors, and unsafe behavior

MCP scans inspect repository contents only. They do not connect to a live MCP endpoint or execute the server.

How scoring works

Every entity is scored across three dimensions: who published it, how it behaves, and what's actually in it.

Each check returns a score from 0–100, a verdict of safe, caution, suspicious, or dangerous, and a confidence level. Cached results are reused across organizations.

New lookups return a preliminary identity score immediately and finish the deep scan in the background. Pass mode=full when you need the completed result in one request.

See the page for each entity for the exact tiers, limits, and threat types.

Use it

  • Dashboard: open Agents in the sidebar, then Context. Submit a web page, raw email, file URL, skill target, public GitHub MCP repository, or package coordinate and open a result to see subscores, tiers, and threats.
  • REST API: see the Context Guardrails API.
  • MCP: get_context_score for pages, scan_email for raw messages, scan_file for file URLs, scan_skill for agent skills, scan_mcp for public GitHub MCP repositories, and scan_package for registry packages.

Next steps