MCP connector

AgentAvow Docs · agentavow.com

AgentAvow's remote MCP server lets Claude (and any MCP client) check whether a tool is safe before an agent connects to it. This page is for users and IT admins evaluating the connector.

What it is

• URL: https://agentavow.com/mcp (remote, Streamable HTTP)
• Auth: none. Every tool is anonymous and read-only.
• What it does: scans a target and returns one of three answers with its reason

(Safe to connect, Review before you connect, or Do not connect), a signed 0–100 trust score with the findings behind it, and an adoption score from real usage (downloads, stars, installs). The scored verdict is signed (Ed25519) and can be recomputed offline.

Tools (all read-only)

• scan_repo — scan a public GitHub repo (owner/name).
• scan_package — scan a published npm / PyPI / crates / Docker / Hugging Face package.

Pin a release inside the name to scan exactly what you would install: chalk@5.3.0 or @scope/name@1.2.3 (npm), requests==2.32.5 (PyPI), serde@1.0.200 (crates). A bare name scans the latest release; a range such as ^1 is ignored and resolves to latest. Advisories against the package are reported for the version scanned, with the fixed release named.

• scan_mcp_server — scan a live MCP server's tool definitions for poisoning / injection.
• verify_trust, check_interaction_safety, lookup_identity, get_trust_badge

— resolve and check an agent's identity and trust before delegating to it.

• about_agentavow — an overview of the service.

Two prompts are also available: agentavow_get_started (what AgentAvow does and how to start) and agentavow_check_my_connections (check the other connectors in this conversation).

None of the tools write, delete, install, or transact. They only read public information and call AgentAvow's own API.

Data handling

• Checking a tool is anonymous — no account, no login, no personal data collected.
• The connector receives only the target you give it (a public repo, package name, or MCP

URL). Scan results about public tools are themselves public.

• **No user content is used for model training. No conversation data or Claude memory is

accessed.**

• Full policy: Privacy · Terms.

Verifying a result

Every verdict is an Ed25519 JWS over an RFC 8785 (JCS) canonical form. Anyone can recompute the verdict byte-for-byte and check the signature offline against the public JWKS — you don't have to trust AgentAvow. See Verify an attestation.

Add it

• Claude Desktop / claude.ai: add it from the

connector directory, or go to Settings → Connectors → Add custom connector → https://agentavow.com/mcp (no authentication).

• Claude Code: claude mcp add --transport http agentavow https://agentavow.com/mcp,

or install the AgentAvow Trust plugin, which adds a /scan command and scans each new MCP server at session start. See Auto-scan in Claude Code.

• Cursor: Add to Cursor
• VS Code: Install in VS Code

Support

Questions or issues: support@agentavow.com.