Blog / Topic
Tools and MCP
Start here
All articles
What should you do with MCP servers that have no owner?
An ownership-claim workflow for discovered MCP servers: keep them in a shadow state, set a claim deadline, block on clear criteria, and record evidence.
Should MCP access be authorised per server or per tool?
Server-level MCP access decides who reaches a server; tool-level authorisation decides which actions they can take. What each stops, and interim controls.
Should MCP servers run as sidecars or shared services on Kubernetes?
Sidecar vs shared-service MCP servers on Kubernetes: how each affects isolation, identity, scaling and audit, and when to pick which.
Which agent APIs let you require approval before an MCP tool runs?
OpenAI Responses and Microsoft Foundry Agent Service support require_approval for remote MCP tools; Anthropic's MCP connector has allow/deny lists only. Compared.
Should AI agents be allowed to search and fetch from the web?
How to govern web search and web fetch tools for AI agents: domain allowlists, use limits, URL exfiltration risk, and what to log for each fetch.
What are MCP Apps and what new risks do they add?
MCP Apps let MCP servers return interactive HTML interfaces in sandboxed iframes. What the extension specifies, and the governance questions it raises.
Should you use official vendor MCP servers or community-built ones?
Official vendor MCP servers usually win on auth, maintenance and accountability. How to verify which is which, and when a community server is acceptable.
How do you verify an A2A agent card before trusting a remote agent?
A practical checklist for verifying A2A agent cards: well-known location, signed cards (JWS over JCS), declared auth schemes, extended cards and card risks.
How should MCP servers store API keys and secrets?
MCP configs often hold plaintext tokens. Where secrets end up, why OAuth beats static keys, secret-manager patterns, and how to scan configs for leaks.
How do you deprecate an MCP server without breaking agents?
A step-by-step way to retire an internal MCP server: find every dependent agent from usage data, set notice periods, use shims, and revoke access last.
Should read and write tools live on separate MCP servers?
Trade-offs of splitting read and write MCP tools across servers versus read-only modes, toolsets and annotations, with a decision table for least privilege.
How do you govern long-running MCP tasks?
What the MCP Tasks extension specifies for long-running tool calls—task handles, TTLs, cancellation, auth binding—and the controls enterprises should add.
What is the difference between function calling and MCP?
Function calling is how a model asks for a tool; MCP is how tools are discovered, transported and authorised. A layer-by-layer comparison and what to govern.
What can real MCP security incidents teach enterprises?
Three documented MCP incidents: a malicious npm server, an mcp-remote RCE and an MCP Inspector flaw. What happened, and the controls each one argues for.
What are Client ID Metadata Documents in MCP authorization?
CIMD lets an MCP client use an HTTPS URL as its OAuth client_id. How the flow works, how it compares with Dynamic Client Registration, and why it helps allow-listing.
Why does an AI agent only see the first page of MCP results?
MCP paginates list operations, not tool results. Why agents treat page one as the whole answer, and how to design and test MCP tools that return lists.
How do you govern browser automation MCP servers like Playwright MCP?
Playwright MCP gives agents a real browser. Settings to restrict origins, isolate profiles, hide secrets and log sessions, plus a policy for developer use.
What are MCP output schemas, and why do they matter for governance?
MCP output schemas let tools declare the shape of their results. What the spec requires, how clients validate structuredContent, and governance uses.
Is it safe to auto-generate MCP servers from OpenAPI specs?
Auto-converting OpenAPI specs to MCP servers risks tool explosion, exposed destructive endpoints, weak descriptions and auth passthrough. A curation checklist.
Can you manage one MCP server configuration across Copilot, Claude Code and Cursor?
Each coding agent reads its own MCP config file. How to run one approved MCP catalogue across GitHub Copilot, Claude Code and Cursor without drift.
Can developers query the AI agent inventory from inside GitHub Copilot or Claude Code through MCP?
How to expose an AI agent inventory to GitHub Copilot and Claude Code as a read-only MCP server, with config examples and the policy switches involved.
What is the confused deputy problem in MCP?
How an MCP proxy with a static client ID and a consent cookie can hand an attacker an authorization code, and the per-client consent fix the spec requires.
Should agents call internal APIs through MCP or OpenAPI tools?
MCP vs OpenAPI tools for internal APIs: how they compare on discovery, auth, schema reuse, governance hooks and client support, plus a rule for choosing.
What changed in the July 2026 MCP specification for enterprise security?
MCP 2026-07-28 removes sessions, deprecates Dynamic Client Registration for CIMD, requires iss validation and binds credentials to their issuer.
What happens when two MCP servers expose tools with the same name?
MCP tool names are only unique per server. How collisions and cross-server tool shadowing happen, how clients namespace tools, and registry rules to stop it.
How do you govern AI agents that control a computer or browser?
Run computer-use agents in isolated low-privilege VMs, allowlist domains, keep credentials out, retain screenshots deliberately and confirm high-impact actions.
How many tools should one AI agent have?
Vendor guidance says fewer than 20 tools up front, with accuracy dropping past 30–50. How to cut the list with tool search, per-task tool sets and sub-agents.
Can you trust MCP tool annotations like readOnlyHint?
MCP tool annotations are hints the spec says to treat as untrusted unless the server is trusted. How to use them in policy without relying on them.
Do MCP roots restrict which files a server can read?
No. MCP roots are informational guidance, not access control, and are deprecated as of 2026-07-28. Restrict file access with sandboxing and OS permissions.
How do you govern MCP resources and prompts, not only tools?
MCP servers expose tools, resources and prompts. Each carries a different risk. A policy per primitive and the inventory fields to track them.
What should you log for every MCP tool call?
A 14-field checklist for MCP tool call logs mapped to OpenTelemetry MCP and GenAI conventions, plus redaction rules for arguments and results.
What is the difference between agent skills and MCP servers?
Skills package instructions and scripts an agent loads on demand; MCP servers give it live tool connections. Risk profiles and a governance table for both.
How do you check agent skills for malicious instructions before use?
A 14-point review checklist for SKILL.md files and bundled scripts: hidden instructions, setup steps, network calls, credential access and provenance.
How do you govern code execution tools given to AI agents?
Sandbox rules for agent code tools: no network by default, CPU, memory and time caps, ephemeral files, hosted vs self-hosted interpreters, and what to log.
How do you check open-weight models before using them in agents?
A pre-use checklist for open-weight models: safe file formats, hub scan limits, remote code, licence terms, hash pinning and behavioural evaluation.
What is the security checklist for hosting a remote MCP server?
A 14-point checklist for a remote MCP server: OAuth per the MCP spec, audience validation, no token passthrough, Origin checks, rate limits and logging.
How do you run a private MCP registry for your enterprise?
Build a private MCP subregistry: mirror approved servers from the official registry, add approval fields, pin versions and point clients at it.
What are MCP sampling and elicitation, and are they safe?
MCP sampling lets servers request model completions; elicitation lets them ask users for input. How each works, the risks, and client-side controls.
What is the difference between MCP and A2A?
MCP connects an agent to tools and data; A2A lets one agent delegate work to another. Compare purpose, discovery, transports, auth and what to govern.
How do you monitor MCP tool call failures across AI agents?
Log every MCP tool call with server, tool, argument hash, latency and error class; separate JSON-RPC errors from isError results; set per-server alerts.
Should you expose internal AI agents as MCP servers, A2A agents or REST APIs?
Choose MCP, A2A or REST for internal agents by consumer type, and see what each means for authentication, discovery, logging and versioning behind an API gateway.
How do you govern an MCP server that turns natural language into SQL?
Give text-to-SQL MCP servers a read-only database role, timeouts, row-level security and full query logging. Treat tool annotations as hints, not controls.
What requirements should an enterprise set for managing AI agent skills?
A requirements list for AI agent skills management in six lifecycle areas, from inventory to retirement, with an acceptance test for each requirement.
How do you govern agent-to-agent (A2A) calls?
Turn A2A Agent Cards, security schemes and signatures into per-agent call policies: allow-lists, delegated identity, hop limits and audit logs.
How does MCP authorization work?
MCP authorization uses OAuth 2.1, Protected Resource Metadata and resource indicators. The flow step by step, what changed in 2026-07-28, and enterprise implications.
What is MCP tool poisoning and how do you prevent it?
MCP tool poisoning hides instructions in tool metadata. The description, rug-pull and shadowing variants, with a control and a ten-point prevention checklist.
Should you allow MCP servers from public registries?
Public MCP registries verify publishers, not code safety. A twelve-point approval checklist and a six-step internal allow-list process for enterprises.
What is an MCP gateway?
An MCP gateway proxies agent tool calls to MCP servers: client auth, per-server credentials, tool filtering, logging and rate limits. How it differs from LLM gateways.
How do you find local MCP servers on developer laptops?
Where Claude, Cursor, VS Code, Windsurf, Gemini CLI and Codex store MCP configs, and a six-step MDM approach to collect and review local MCP servers.
How do you stop prompt injection that arrives through tool outputs?
Nine defences against indirect prompt injection via tool outputs, ranked by effectiveness, with what each cannot stop and a seven-step rollout.
What should you record for every MCP server in your inventory?
A 15-field MCP server inventory checklist: owner, transport, auth method, pinned version, tool definition hashes, data scopes, consuming agents and review dates.
Is a local stdio MCP server safer than a remote HTTP one?
stdio vs Streamable HTTP MCP servers compared: network exposure, DNS rebinding, process privileges, authentication and a security checklist for each transport.
How do you sandbox an MCP server?
Ten steps to sandbox MCP servers: containers, non-root, read-only filesystems, minimal mounts, egress allow-lists, per-server credentials and resource limits.
How do you limit which tools each AI agent can call?
Deny by default, allow-list tools per agent version, constrain arguments and enforce outside the prompt. Worked examples and a policy checklist.
How do you review MCP server updates before agents use them?
Pin versions, hash tool definitions, diff every update and define what forces re-approval. A practical review workflow for local and remote MCP servers.
What is agent skills governance?
Agent skills governance defined: controls over the SKILL.md packages agents load, across authoring, review, versioning, distribution, revocation and usage.
How do you review MCP servers your agents already use?
A ten-point review checklist for MCP servers you have discovered: publisher, version pinning, transport, auth, tool descriptions, write tools, data reach and owner.
Put the framework into practice