VeriSwarm Tools MCP

The Tools MCP at https://mcp.veriswarm.ai is what an AI agent — or a human's AI assistant — calls while it works: scan a tool definition before trusting it, check whether an action is allowed, tokenize PII before it leaves your boundary, or hand a task to another agent. It is one of several MCP surfaces; this is the runtime one.

Connecting

The Tools MCP uses OAuth 2.0. There are two paths:

• A human, via an AI assistant (Claude or ChatGPT connectors): add https://mcp.veriswarm.ai as a custom connector. Your assistant runs the standard authorization_code + PKCE consent flow; you log in and approve the scopes.
• An agent, machine-to-machine: exchange a VeriSwarm agent API key for a token via client_credentials against https://mcp.veriswarm.ai/oauth/token, then send it as a Bearer token.

Discovery documents:

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

Scopes

Four scopes gate the tools. A connector that registers without naming scopes receives only guard:scan and trust:check — pii:tokenize and a2a:invoke must be requested explicitly.

Scope	Grants	Plan required
guard:scan	security scanning tools	Guard (Max)
trust:check	trust-decision and score reads	— (free tier)
pii:tokenize	PII tokenization	Guard (Max)
a2a:invoke	agent-to-agent catalog + task submission	Pro or above

Tools

Fourteen tools, grouped by scope. whoami needs no scope.

guard:scan

• scan_agent_definition — scan an agent's definition/SOUL for risks
• scan_mcp_tools — scan MCP tool definitions (the 10-check scanner)
• scan_session_turn — score one conversation turn for exfiltration risk
• validate_tool_schema — validate a tool's JSON schema
• check_tool_permission — check whether a tool call is permitted

trust:check

• check_decision — evaluate an allow/review/deny decision for an action
• get_agent_trust — read an agent's trust score and factors

pii:tokenize

• tokenize_pii — replace PII with reversible tokens before sending onward
• list_pii_session_tokens — list tokens issued in a session

a2a:invoke

• list_a2a_agents — trust-ranked catalog of agents
• get_a2a_agent_card — an agent's capability card (with x-veriswarm-trust)
• submit_a2a_task — submit a task to another agent
• get_a2a_task_status — poll a submitted task

No scope

• whoami — report the caller's identity, tenant, and granted scopes

PII rehydration is not available here

tokenize_pii replaces PII with tokens, but restoring the original values is deliberately not offered on this hosted surface. Rehydration lives only on the self-hosted server, which runs inside your own trust boundary. This is a security boundary, not an omission.