MCP security, governance & how-tos
Practical guides for anyone connecting AI to real tools — from the team building gate, the MCP gateway.
MCP Ping Explained: The Protocol's Smallest Request
Either side can send a bare ping and expect an empty reply back — MCP's minimal liveness check, with no parameters, no capability negotiation, and no opinion on whether a tool call would actually succeed. What it's for, what it doesn't prove, and where it quietly matters for remote connections.
MCP Cancellation Explained: What Actually Happens When You Stop a Tool Call
notifications/cancelled lets either side ask the other to abandon an in-flight request — but it's best-effort, can arrive too late, and can't undo a side effect a tool already triggered. How it's shaped, what it doesn't guarantee, and where it actually matters.
MCP Structured Output Explained: How a Tool Result Can Carry Real JSON, Not Just Text
A tool can declare an outputSchema and return structuredContent alongside the usual text block, instead of leaving a model to parse JSON back out of a string. How it's shaped, what it doesn't guarantee, and where it actually matters.
MCP Resource Subscriptions Explained: How a Client Learns a Resource Changed
A client can ask a server to push notifications/resources/updated instead of re-reading a resource on a timer — a narrower, quieter mechanism than resources/list_changed, and one most servers never implement. How it's shaped, what it doesn't guarantee, and where it actually matters.
How to Connect Perspective to Claude with MCP
Perspective's remote MCP server puts mobile funnels, leads, and campaign stats inside Claude — but its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and what to check before an AI client reads a real lead list.
How to Connect iLook Face Analysis to Claude with MCP
iLook's remote MCP server scores face shape, symmetry, and Golden Ratio proportions from a photo, with no sign-in and no identity recognition. Here's how to connect it directly or through a gateway, and why a tool built around a real face deserves a different kind of care than one that only reads text.
How to Connect Higgsfield to Claude with MCP
Higgsfield's remote MCP server turns a prompt into a cinematic AI image or video right inside Claude, and its OAuth just works with a new client. Here's how to connect it directly or through a gateway, and why a generation tool deserves review after it produces something, not instead of it.
How to Connect Squarespace to Claude with MCP
Squarespace's remote MCP server doesn't touch your site, store, or billing — it only searches and generates available domain names, and it needs no sign-in to use. Here's how to connect it directly or through a gateway, and why a server's brand name isn't its scope.
MCP Tool Name Collisions: What Happens When Two Servers Share a Tool Name
Connect two MCP servers that both define a tool called search, and the protocol has no required answer for what happens next. What a naming collision actually looks like, how it differs from tool shadowing, and how a gateway avoids it by namespacing every tool by server instead of detecting conflicts after the fact.
How to Connect Jithox E-Invoice to Claude with MCP
Jithox's remote MCP server preflights EU e-invoicing before you send it: invoice structure, live VAT lookups via VIES, and Peppol reachability — no sign-in required. Here's how to connect it directly or through a gateway, and why a validation tool still deserves care with real invoice data.
PII in MCP Tool Results: What Filtering Can Catch, What It Can't, and What to Do First
Every MCP tool result lands in the model's context. PII filtering on a gateway can redact emails and IBANs, but misses context, gets reassembled, or breaks the tool. Why the first control is the tool surface, not the regex, and a checklist for both.
Data Exfiltration Through MCP: The Four Paths Out and How to Close Them
Once an AI client is tricked, how does the data actually leave? A write tool on another server, a URL argument on a benign tool, the server operator itself, or the model's own retention. Each path, and the control that closes it.
Private MCP Registry: Why an Allow-List Beats a Blocklist for AI Tools
The public registry lists tens of thousands of servers; your company needs twelve. What a private MCP registry actually is, what each entry should carry, why it has to be enforced at a gateway, and how to make the request path fast enough that people use it.
GDPR and MCP: What Happens to Personal Data When Your AI Calls a Tool
An MCP tool call moves customer data through an AI client into a third-party model. Who is the processor, what Article 28, 30 and 32 expect, where transfers happen, and a practical sequence for controllers. Not legal advice; the map your counsel wants drawn.
HIPAA and MCP: Connecting AI to Systems That Hold PHI
The moment an MCP tool returns a patient name next to a diagnosis code, HIPAA applies to every hop. Business associate agreements along the chain, minimum necessary as tool rules, the Security Rule's technical safeguards, and the log you'll need for breach scope.
DORA and MCP: AI Tool Access as ICT Third-Party Risk
DORA has applied since January 2025 and makes no exception for the analyst who connects an MCP server to Claude. The five pillars read against an MCP deployment: register of information, risk management, incident reporting, testing, and third-party due diligence.
Anomaly Detection for AI Tool Calls: What "Abnormal" Looks Like in an MCP Log
Agents are bursty, erratic, and the attacker never logs in, so rate and identity alerts fail. The signals that actually mean something in MCP traffic: read-then-send across servers, secrets in arguments, new external hosts, drift, destructive bursts. Rules over a classified log, not a model.
How to Connect Hugging Face to Claude with MCP
Hugging Face's remote MCP server lets Claude search models, datasets, and Spaces on the Hub directly, with no sign-in required. Here's how to connect it directly or through a gateway, and why a popular model card still deserves its own read.
How to Connect DeepWiki to Claude with MCP
DeepWiki's remote MCP server turns any public GitHub repo into a wiki Claude can read and ask questions against — no sign-in, no cloning, just three tools for structure, contents, and direct Q&A. Here's how to connect it directly or through a gateway, and why a generated wiki is a summary, not the source.
How to Connect Context7 to Claude with MCP
Context7's remote MCP server resolves a library name to its current documentation and code examples, so Claude stops recommending an API that changed or was deprecated since training. Here's how to connect it directly or through a gateway, and why 'no sign-in' still deserves a quick check.
How to Connect Apify to Claude with MCP
Apify's remote MCP server puts thousands of scraping and automation actors inside Claude — and unlike most servers in this series, its listing says OAuth but the verified connect path resolves through an API key. Here's how to connect it directly or through a gateway, and why running someone else's actor is a different kind of trust than calling a fixed tool.
MCP Logging Explained: How a Server Streams Diagnostic Messages to the Client
MCP has a logging/setLevel request and a notifications/message stream for free-text diagnostics — a different channel than a tool result's isError, an audit log, or progress notifications. What it's for, how the severity levels work, and why almost no client surfaces it.
How to Connect Square to Claude with MCP
Square's remote MCP server puts payments, orders, catalog, and customers inside Claude — but unlike Stripe and PayPal, its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and what to check before an AI client touches a live business account.
How to Connect Spotify to Claude with MCP
Spotify's remote MCP server is real and official, but gate's directory tracks its endpoint as an external pilot with restricted connectivity — not a general release. Here's what it does, how to attempt a connection directly or through a gateway, and why a pilot wall is a real possibility, not a broken setup.
How to Design MCP Tools Around Agent Jobs, Not API Endpoints
Every generator defaults to one MCP tool per API operation, because that's the mechanical part. Whether that's the right shape for an agent is a separate question. When to combine steps into one job-shaped tool, when to keep them atomic, and what each choice costs in flexibility, retries, and error handling.
How to Connect X (Twitter) to Claude with MCP
X's official remote MCP server lets Claude search and post on your account — but it has no RFC 9728 metadata for a client to auto-discover OAuth, so you connect with a bearer token instead of a one-click consent screen. Here's how to connect it directly or through a gateway, and why a posting tool deserves a different kind of caution than one that only reads.
How to Connect Box to Claude with MCP
Box's remote MCP server puts document and folder search inside Claude — but its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and what to check before an AI client can read or move a file your whole team shares.
Excessive Agency in MCP: When a Tool Can Do More Than the Task Needs
A tool can be exactly what it claims to be and still carry more capability than any one task requires — a generic action field, a write-capable credential behind a read-looking tool. Why that's different from OAuth scope creep and tool annotations, how to spot it before you connect a server, and how to design against it.
How to Connect Typeform to Claude with MCP
Typeform's remote MCP server puts forms, responses, and insights inside Claude — but its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and why form responses deserve a slower look than most data in this series.
MCP Progress Notifications Explained: How a Server Reports Status on a Slow Call
A request can carry a progressToken that lets a server stream small status updates back while the call is still open — a different fix than MCP Tasks, and one that predates it. How the notification is shaped, what the one hard rule on progress values means, and why it's not a substitute for a real progress bar.
How to Connect Userpilot to Claude with MCP
Userpilot's remote MCP server puts onboarding flows, NPS, and product usage data inside Claude, and its OAuth just works with a new client. Here's how to connect it directly or through a gateway, and what to check before an AI client reads free-text survey comments from real users.
How to Paginate Large MCP Tool Results (When the Protocol Doesn't Do It for You)
MCP has cursor-based pagination for listing a server's tools, resources, and prompts — but nothing built in for a single tool call that returns a huge result. The patterns that keep a search or list tool from blowing up the model's context, and the two that don't work.
How to Connect Amplitude to Claude with MCP
Amplitude's remote MCP server puts events, charts, and cohorts inside Claude, and its OAuth just works with a new client. Here's how to connect it directly or through a gateway, and what to check before an AI client reads your product's event data.
MCP Completion Explained: How Autocomplete Works for Prompt and Resource Arguments
The completion/complete request lets an MCP server suggest values as a client types a prompt argument or a resource URI — a narrow, optional primitive that has nothing to do with tool calls. What it covers, how the request and response are shaped, and why almost no server bothers to implement it.
How to Connect Fathom to Claude with MCP
Fathom's remote MCP server puts meeting transcripts, summaries, and action items inside Claude, and its OAuth just works with a new client. Here's how to connect it directly or through a gateway, and why a shared meeting archive deserves a slower look than tools that only expose your own data.
How to Handle Errors in an MCP Tool So the Model Can Actually Recover
MCP gives tool authors a specific channel for a failed operation that isn't a protocol error: a normal result with isError set to true. Why that distinction matters, what makes an error message useful to a model instead of just a human, and the mistakes that turn a loud failure into a silent, wrong one.
How to Connect Gamma to Claude with MCP
Gamma's remote MCP server turns a prompt into a presentation, document, or webpage inside Claude — but its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and why a generation tool deserves review after it produces something, not instead of it.
How to Connect Canva to Claude with MCP
Canva's remote MCP server puts design generation and editing inside Claude — but its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and why a tool that creates new content deserves a different check than one that only reads or edits.
How to Connect Strava to Claude with MCP
Strava's remote MCP server puts activities, stats, and training data inside Claude — but its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and why GPS activity data deserves a slower look than most tools in this series.
How to Debug an MCP Server with the MCP Inspector
The MCP Inspector shows the raw protocol traffic a chat interface hides — the handshake, the exact tool list, and the request and response for every call. How to run it, what to look at, and what it won't catch on its own.
How to Connect Miro to Claude with MCP
Miro's remote MCP server puts board reading and diagramming inside Claude, and its OAuth just works with a new client. Here's how to connect it directly or through a gateway, and what to check before an AI client can add to or rearrange a board your whole team shares.
How to Connect ElevenLabs to Claude with MCP
ElevenLabs' remote MCP server puts text-to-speech and voice generation inside Claude — but its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and why a tool that can produce a voice deserves a different kind of caution than one that reads or writes text.
MCP Server Versioning: How to Change a Tool Without Breaking Every Client
MCP has no version negotiation, no deprecation header, and no schema diffing built in — a tool's inputs or meaning can change and the only client that finds out is whichever one calls it next. Patterns that actually work: versioning the tool name, additive-only changes, and testing against the last published shape before you ship.
One MCP Server, Several AI Clients: Do You Need a Separate Connection for Each?
Claude, Cursor, and ChatGPT each keep their own MCP config, so the same server usually means a separate connection, and a separate OAuth grant, in every client. What that costs once you're past one client and one server, and how a gateway collapses it into one connection.
How to Connect Firecrawl to Claude with MCP
Firecrawl's remote MCP server turns scraping, searching, and parsing the web into tools Claude can call directly, and it needs no sign-in to connect. Here's how to connect it directly or through a gateway, and why scraped content deserves the same caution as any other untrusted input.
What to Do When an MCP Server You Use Might Be Compromised
A suspicion about a compromised MCP server is a reason to revoke first and investigate second, not the other way around. The sequence: cut its access, check what it actually called, rotate what its token could reach, re-scan, and decide whether it comes back.
How to Connect Qonto to Claude with MCP
Qonto's remote MCP server puts business account balances, transactions, and invoices inside Claude, and its OAuth just works with a new client. Here's how to connect it directly or through a gateway, and why a live bank account deserves a slower look than most tools in this series.
How to Test an MCP Server Before You Publish It
A clean security scan tells you a tool's descriptions aren't hostile — it doesn't tell you they're clear, that your annotations are accurate, or that a retry won't double-charge someone. A practical pre-publish checklist for MCP server builders.
How to Connect Dropbox to Claude with MCP
Dropbox's remote MCP server puts file search, reading, and organizing inside Claude — but its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and what to check before an AI client can move or delete a shared file.
MCP Elicitation Explained: How a Server Can Pause and Ask You Something Mid-Task
Elicitation lets an MCP server pause mid-tool-call and ask a person a direct question through a simple form, instead of routing it through the model as free text. What the elicitation/create request looks like, how it differs from sampling, and what the spec bans it from asking for.
MCP Tasks Explained: How Long-Running Tool Calls Work
MCP Tasks, one of the extensions in the July 2026 spec update, gives servers a standard start-check-fetch pattern for slow work instead of a private polling scheme. What it changes, why it pairs with MCP going stateless, and when it's actually worth using.
How to Connect Intercom to Claude with MCP
Intercom's remote MCP server puts conversations, contacts, and help content inside Claude — but its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and what to check before an AI client can reply to a real customer.
How to Connect Wix to Claude with MCP
Wix's remote MCP server puts site content, bookings, and products inside Claude, and its OAuth just works with a new client. Here's how to connect it directly or through a gateway, and what to check before an AI client can cancel a booking or change a product's price.
MCP Roots Explained: How a Client Tells a Server Which Directories It Can Touch
Roots are the quiet fourth MCP primitive: a list of URIs a client hands a server marking out what's relevant to touch. What roots actually do, why they're advisory rather than enforced, and what that means if you're building or connecting a filesystem-facing server.
How to Connect ClickUp to Claude with MCP
ClickUp's remote MCP server puts tasks, docs, and spaces inside Claude — but its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and what to check before an AI client can create or reassign a task.
How to Connect Asana to Claude with MCP
Asana's remote MCP server puts tasks, projects, and portfolios inside Claude — but its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and what to check before an AI client can create or reassign a task.
MCP Tool Annotations Explained: readOnlyHint, destructiveHint, and Why They're Only Hints
MCP tool definitions can carry readOnlyHint, destructiveHint, and two other annotations describing a tool's own risk. Almost no server sets them, and the spec is explicit they're not guarantees. What they are, why gate treats them as one signal among several, and what to set if you're building a server.
How to Connect Todoist to Claude with MCP
Todoist's remote MCP server puts your tasks and projects inside Claude, and its OAuth just works with a new client. Here's how to connect it directly or through a gateway, and what to check before an AI client can create or clear a shared project.
MCP Server Supply Chain Risk: What You're Trusting When You Install One
Running a local MCP server with npx or uvx means executing someone else's code on your machine with your full user permissions, before a single tool call happens. What that risk actually covers, why it's separate from a tool-description scan, and what to check before you run one.
How to Connect PayPal to Claude with MCP
PayPal's remote MCP server puts invoices, orders, and transactions inside Claude, and its OAuth just works with a new client. Here's how to connect it directly or through a gateway, and what to check before an AI client can create an invoice or trace a payment on your behalf.
How to Connect Neon to Claude with MCP
Neon's remote MCP server puts Postgres project management, branching, and querying inside Claude, and its OAuth just works with a new client. Here's how to connect it directly or through a gateway, and what to check before an AI client can run SQL against a real database.
MCP Tool Idempotency: Why Retries Can Silently Duplicate Actions
An AI client retrying a stalled tool call is normal behavior, not a bug — but a write tool that isn't idempotent turns that retry into a duplicate email, ticket, or charge. What idempotency means for MCP tools, why retries happen, and how to design tools that survive being called twice.
What an MCP Security Scanner Actually Checks (and What It Can't)
A scan result reads like a verdict, but it's really three narrow checks on tool text: pattern matching, an LLM review pass, and provenance. What gate's scanner catches, what drift detection adds, and the failure modes no text-based scan can see.
How to Connect DocuSign to Claude with MCP
DocuSign's remote MCP server puts envelope status, signers, and sending inside Claude — but its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and why a tool that can send an envelope deserves more caution than one that only reads.
MCP Apps Explained: What It Means for a Tool Result to Render as UI
MCP Apps, one of the extensions in the July 2026 spec update, lets a server return real interface — forms, pickers, confirmations — instead of text the model has to describe. What it changes, what it doesn't, and why it's early.
How to Revoke an MCP OAuth Grant (And What Revoking Doesn't Always Do)
A grant you approved once can live in three places at once: the vendor's OAuth settings, your AI client's connection list, and a gateway's own record. Where to revoke it, in what order, and how to check it actually worked.
How to Connect Webflow to Claude with MCP
Webflow's remote MCP server puts sites, pages, and CMS content inside Claude, and its OAuth just works with a new client. Here's how to connect it directly or through a gateway, and what to check before an AI client can edit a live site.
How to Connect Cloudflare Docs to Claude with MCP
Cloudflare's documentation MCP server skips OAuth entirely — no account, no sign-in, nothing to authorize. Here's what it does, how to connect it directly or through a gateway, and why 'no sign-in' still deserves a quick check before you trust it.
MCP Sampling Explained: How a Server Can Ask the Client's Model a Question
MCP defines a fourth primitive beyond tools, resources, and prompts: sampling, where a server asks the client's own model for a completion mid-task. What it's for, why the client stays in control, and why almost no server uses it yet.
How to Connect PostHog to Claude with MCP
PostHog's remote MCP server puts your analytics, insights, and feature flags inside Claude — and unlike several others in the catalog, its OAuth just works with a new client. Here's how to connect it directly or through a gateway, and what to check before an AI client reads your product data.
How to Connect Figma to Claude with MCP
Figma's remote MCP server puts design files, frames, and specs inside Claude — but its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and what to check before an AI client can read a design file.
MCP Token Passthrough: The Anti-Pattern the Spec Warns Against
Some MCP servers forward the caller's own access token straight to their backend instead of issuing an audience-scoped one. Why that shortcut is tempting to build, why it quietly breaks OAuth's safety guarantees, and how to tell if a server does it.
How to Connect Vercel to Claude with MCP
Vercel's remote MCP server puts deployments, projects, and logs inside Claude — but its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and what to check before an AI client can read your build logs.
How to Add OAuth to Your Own MCP Server
A remote MCP server needs real OAuth 2.1 — PKCE, dynamic client registration, resource-scoped tokens — not an API key stapled to a header. What a working implementation needs, where builders get stuck, and the shortcut that skips most of it.
How to Connect Zapier to Claude with MCP
Zapier's remote MCP server puts every Zap you've built one prompt away from firing — which means it inherits the reach of every app those Zaps touch. Here's how to connect it directly or through a gateway, and what the wider scope means for what you approve.
MCP Tool Shadowing: When One Server's Description Rewrites How Another Behaves
A tool description on one MCP server can contain text that changes how the model calls a tool on a completely different, fully legitimate server. What tool shadowing is, why per-server review misses it, and what actually limits it.
How to Connect Airtable to Claude with MCP
Airtable's remote MCP server puts your bases, tables, and records inside Claude, and its OAuth just works with a new client. Here's how to connect it directly or through a gateway, and what to check before an AI client can read or edit a shared base.
MCP Tools, Resources, and Prompts: What's the Difference?
Most MCP servers only ever implement tools. The spec also defines resources and prompts — built for the cases tools handle badly. Who controls each one, when to use which, and why builders default to tools anyway.
How to Connect Supabase to Claude with MCP
Supabase's remote MCP server lets Claude query your database, inspect your schema, and manage a project directly. Here's how to connect it directly or through a gateway, and why the project you point it at matters more than usual.
The Confused Deputy Problem in MCP: When Your AI Uses Its Own Access Against You
A confused deputy attack doesn't need a malicious server or a broken OAuth flow — just an agent with real access and one piece of untrusted content. What the classic security problem looks like in MCP tool chains, and what actually limits it.
How Many Tools Should an MCP Server Expose?
MCP doesn't cap tool count, but the model does. Why a long tool list quietly hurts selection accuracy, where the number usually comes from, and how to curate a server down instead of mirroring an entire API.
How to Connect HubSpot to Claude with MCP
HubSpot's remote MCP server puts contacts, deals, and marketing data inside Claude — but its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and what to check before an AI client can read or edit your CRM.
MCP "Unauthorized Client" Error: Why It Happens and How to Fix It
You add an MCP server and OAuth rejects your client with no clear reason. Here's why some vendors allow-list clients instead of open registration, how common it actually is, and three ways to connect anyway.
How to Connect Slack to Claude with MCP
Slack's remote MCP server puts channels, threads, and messages inside Claude — but its OAuth currently only admits allow-listed clients. Here's how to connect it directly or through a gateway, and what that restriction means in practice.
How to Tell a Real MCP Server from a Fake One
Nothing stops anyone from publishing an MCP server called "Notion" or "Slack." Here's what actually verifies a server is what it claims to be — the domain, the vendor's own docs, and the OAuth redirect — before you connect it.
How to Write MCP Tool Descriptions Your AI Will Actually Use Correctly
A tool description is the only spec a model sees before it calls your tool. Here's how to name tools, what to put in the description versus the schema, and how to avoid writing one that doubles as a prompt-injection vector.
How to Connect Stripe to Claude with MCP
Stripe's remote MCP server puts customers, payments, and subscriptions inside Claude. Here's how to connect it directly or through a gateway, and what to check before an AI client can look up — or refund — a real charge.
How to Roll Out MCP to Your Team: A Practical Plan
The fastest way to turn MCP into a security incident is connecting everything on day one and sorting out rules later. A rollout plan that starts with one server, one team, and rules decided in advance instead.
How to Connect monday.com to Claude with MCP
monday.com's remote MCP server puts boards, items, and updates inside Claude. Here's how to connect it directly or through a gateway, and what to check before an AI client can read or edit a shared work board.
What to Look For in an MCP Gateway: A Practical Checklist
Not every product wearing the label 'MCP gateway' does the same job. Five things a real one should do — one URL, per-tool rules, security scanning, centralized OAuth, and a readable log — and how to test a candidate gateway for each.
How to Connect Sentry to Claude with MCP
Sentry's remote MCP server puts issues, releases, and stack traces inside Claude. Here's how to connect it directly or through a gateway, and what to check before an AI client starts reading production error data.
How to Connect Jira and Confluence to Claude with MCP
Atlassian's remote MCP server covers Jira, Confluence, and Bitbucket in one connection. Here's how to connect it to Claude directly or through a gateway, and what a single grant across three products actually means for your access.
We Probed 516 MCP Registry Servers. 94% Are Alive — the Dead-Server Crisis Is Elsewhere
Everyone cites '52% of MCP servers are dead.' We ran real MCP handshakes against 516 remote endpoints from the official registry: 93.8% answer. Both numbers are true — and the difference is where you discover servers, not whether servers rot.
The EU AI Act Reaches Your AI's Action Layer: What MCP Users Need to Log
High-risk obligations become binding August 2, 2026 — and Article 12's automatic logging arguably covers AI tool calls via MCP. What a compliant action-layer log takes, why per-employee configs can't provide one, and the gateway pattern that can.
The OWASP MCP Top 10, Explained for Builders
OWASP now has a dedicated MCP Top 10 of security risks. Here's what all ten (MCP01–MCP10) actually mean in plain terms, grouped by root cause, and honestly which a gateway helps with.
MCP Goes Stateless: What the 2026-07-28 Update Actually Changes
Anthropic just shipped the biggest MCP revision since streamable HTTP: a stateless core, MCP Tasks, MCP Apps, and hardened OAuth. What it means for server builders, users, and gateways — in plain language.
How to Connect GitHub to Claude with MCP
GitHub's remote MCP server doesn't hand out one-click OAuth like most others — here's how to connect it to Claude with a personal access token, directly or through a gateway, and what to scope before you do.
MCP Rug Pulls: When a Server Changes the Tool After You Approved It
A tool you approved can be redefined by its server later, and your AI client won't notice. Here's how MCP rug pulls work, why a one-time review can't catch them, and what drift detection does instead.
What Did Your AI Actually Do? MCP Observability & Audit Logs
When your AI takes real actions, you need a record of them. Here's what an MCP audit log should capture, why you'll need it (debugging, security, EU AI Act), and how to make logs people actually read.
How to Turn Any API Into an MCP Server (From an OpenAPI Spec)
Most apps have an API but no MCP server yet. Here's how to turn a REST API into one — the tools that convert an OpenAPI spec, the tool-count and auth traps, and how to do it safely.
How to Connect Linear to Claude with MCP
Two ways to connect Linear to Claude with the Model Context Protocol — direct or through a gateway — plus what to check before you hand an AI client write access to your issue tracker.
MCP OAuth Scopes Explained: What You're Actually Granting When You Connect a Server
Clicking "Allow" on an MCP OAuth consent screen grants more than most people realize. Here's what OAuth scopes actually control in MCP, where they fall short, and how to close the gap.
Remote vs. Local MCP Servers: Which Should You Use?
Local MCP servers run as a process on your machine over stdio. Remote ones run in the cloud over HTTP with OAuth. Here's the practical difference, and how to pick the right one.
How to Connect Notion to Claude with MCP
Two ways to connect Notion to Claude with the Model Context Protocol — direct or through a gateway — plus what to check before you hand an AI client access to your workspace.
Human-in-the-Loop Approvals for AI Tool Calls: Patterns That Work
Not every AI action should run unsupervised. Here's how human-in-the-loop approvals for tool calls work, which calls to gate, the patterns that scale, and how they quietly fail.
What Is an MCP Gateway? (And When Your Team Needs One)
An MCP gateway is a single control point between your AI clients and every MCP server they use. Here's what it does, the problems it solves, and how to tell when your team needs one.
MCP Prompt Injection and Tool Poisoning: How the Attacks Work (and How to Stop Them)
Prompt injection and tool poisoning are the two attacks that make MCP dangerous. Here's how they work, real 2026 examples, and the layered defenses that actually stop them.
Shadow MCP: The New Shadow IT Hiding in Your AI Stack
Anyone can connect an MCP server in minutes, and most never get security review. Here's what Shadow MCP is, why it's spreading, and how to bring ungoverned AI connections back under control.
MCP Access Control for Teams: RBAC and Least Privilege for AI Tools
Rolling MCP out to a team means deciding who can use which tools. Here's how to apply role-based access control and least privilege to MCP — without shared keys or spreadsheets.
How to Vet an MCP Server Before You Connect It: A Security Checklist
Connecting an MCP server hands it real power over your data and systems. Use this practical checklist to vet any MCP server for prompt injection, permission creep, and credential risks before you connect it.