出荷前に MCP 統合のスコープとリスクを評価。

MCP (Model Context Protocol) is an open standard introduced by Anthropic for connecting LLMs and agents to external tools, data, and systems through a uniform server interface. It is increasingly the default integration layer for production agents.

Because scoping and risk-assessing a real MCP rollout is the bottleneck. You decide which systems to connect, which servers per system, the auth scheme, the gateway question, and the OWASP-mapped risk hotspots. Twelve minutes versus a multi-week scoping engagement.

Local stdio for single-user dev tools and strict on-prem with high sensitivity. Remote SSE/HTTP for cloud deployments at any meaningful scale. Hybrid when you mix low- and high-sensitivity workloads. Gateway when you have writes + scale + regulatory.

When you have high autonomy + write-capable servers, or scale (10K+ MAU) + multi-tenant + a regulatory constraint. The gateway centralises OAuth, rate-limiting, prompt-injection filtering, and audit log.

See the matrix on the methodology page. B2C → OAuth 2.1 + short tokens. B2B with high sensitivity → OAuth 2.1 user delegated + per-session rotation. On-prem regulated → mTLS + short-lived JWT issued by internal CA.

Read-only by default, write only when explicitly required, admin scopes never. Per-user audit log. Allowlist tools per server. Rate limit at the gateway when one is recommended, otherwise at the server.

Yes when scoped correctly: short-lived tokens, mTLS for on-prem, full audit log, per-tool allowlist, and human-in-the-loop on writes. Each regulation adds specific requirements you can find on the methodology page.

User→LLM = LLM01 prompt injection. LLM→MCP = LLM06/LLM08 over-permissive scope and excessive agency. MCP→Downstream = LLM08/LLM09 exfiltration. Downstream→LLM = LLM01/LLM03 indirect injection. LLM→User = LLM02 sensitive disclosure.

Because multipliers compound. A four-server integration at the wrong autonomy + regulatory level easily doubles. We always show a range with the breakdown so you can see what is driving the spread.

Use the official one when it exists. High-quality community servers when not. Build custom only when you have a domain-specific protocol or compliance constraint that no public server meets.

Yes. Almost every server in the registry can run on your own infrastructure. The tool flags which ones, what dependencies they need, and the typical sizing.

For strict residency, host the MCP server in-region or on-prem. The tool surfaces residency implications in the per-system table when your regulatory constraints are set.

modelcontextprotocol.io weekly scan, MCP.so quarterly, PulseMCP quarterly, spec repo RSS in real-time. Editorial review quarterly.

Tool allowlist per server, per-session scope narrowing, output scanning for known injection patterns, and per-tool confirmation on writes when autonomy is below post-review.

Gateway-first, then a phased server rollout (3 servers per quarter is realistic at first). Centralise audit + observability before scaling beyond five servers.