Arclan / Methodology

How Arclan evaluates MCP servers

Every server in the registry is validated by live Model Context Protocol handshake, scored on observed reliability, and monitored continuously. A directory lists servers; Arclan tests them. Methodology version 2026-09-29.

Validation

Each candidate is contacted directly. Arclan opens the MCP initialize handshake, then tools/list, following cursor pagination. A 10-second hard timeout applies to every request. What the server actually returns — protocol version, capabilities, tool schemas, latency — is recorded; nothing is taken from a manifest or self-description.

Server states

activeFull handshake passed; connectable now.
auth_gatedReachable but requires credentials to complete the handshake.
mcp_partialResponds over HTTP but fails full MCP compliance.
degradedWas active, now intermittently failing.
stdioLocal transport only — not network-callable, not remotely validated.
unreachable / invalid / zombieNo HTTP response / not an MCP server / was live, now permanently dark.

Score

The composite score (0–100) is a blend of two rollups:

When both exist the composite is 80% Arclan / 20% agent (flipped to 20% / 80% for auth_gated servers Arclan can't fully probe); when only one source exists, that one is used. Agent reports can never lift the composite more than 15 points above the Arclan baseline, so bulk-submitted reports can't game a score. It is an operational reliability metric, not a user rating.

Production-safe

A server is flagged production-safe only when all of the following hold:

These exact thresholds are what the scoring job enforces (the numbers above are read from the same constant the code uses). It is an observed-reliability bar, not a security judgement — read it alongside the security scan below.

Security scanning

After every successful tools/list, tool names and descriptions are scanned for indicators of: prompt injection, arbitrary execution, exfiltration, system access, and credential harvesting.

Detection method: static_manifest_scan

Limitations: Static analysis of the tool's declared name and description (its manifest) only. Arclan does not observe, intercept, sandbox, or execute the tool at runtime, and does not inspect server source, dependencies, or network traffic. A flag indicates threat-associated language in the declared manifest, not observed malicious behaviour; the absence of flags is not a safety guarantee.

Freshness

Trust metrics (uptime, latency, handshake success) roll over 24-hour and 7-day windows and carry a freshness timestamp. Each server page and API response states when the server was last successfully checked, so a consumer can discount stale data.