A MCP server (MCP server) is the program that allows an application like Claude, ChatGPT, Cursor or VS Code to read data and act in another software via a common protocol, the Model Context Protocol.
This guide explains the principle, then how to install a server and the precautions to take. Each statement comes from the official protocol documentation, d Examples of usage refer to directory records with their source.
See also the library of AI agents' skills and Create an AI Agent Guide and the app guide and all use cases.
The Model Context Protocol, or MCP, is an open and open source standard that is used to connect an AI application to external systems: files, a database, a search tool, a business software. The official documentation compares it to an USB-C port: a common socket, instead of a different cable for each device.
Anthropic I presented and published in open source on November 25, 2024, with a specification, development kits (SDK), support for local servers in the Claude Desktop application and a ready-to-use server repository for Google Drive, Slack, GitHub, Git, Postgres and Puppeteer. The problem, according to the announcement, was that each new source of data required its own integration. According to the same announcement, the protocol was created in Anthropic by David Soria Parra and Justin Spahr-Summers.
The specification says to be inspired by the Language Server Protocol, which has unified support for programming languages in code editors. MCP does the same for the context and the tools provided to the AI applications. The documentation lists among the compatible software Claude, ChatGPT, Visual Studio Code and Cursor. In ChatGPT, MCP servers run the apps: seeChatGPT app guide.
The protocol distinguishes three roles. Lhostis the application of the AI you use, for example Claude Desktop, Claude Code or VS Code. For each server it connects to, host creates aclient MCP, which keeps a dedicated connection with this server. The MCP serveris the programme that provides context and actions; it can run on your computer ("local" server) or on the Internet ("remote" server).
A server offers three kinds of elements, which the documentation calls primitives:
The client discovers what the server proposes by list queries (tools/listfor tools), then call a tool withtools/call. Messages follow JSON-RPC 2.0 format. Two transports are planned:stdio, when the server is a program run on the same machine, and Streamable HTTP, for a remote server that can serve many clients. For the latter, the documentation recommends OAuth to get access chips.
The dialogue also goes in the other direction: by "licensing", a server can request clarification or confirmation from the user. Documentation prevents you from asking for a password, a key to the password or payment data by form; These exchanges go through a separate open page (URL mode), out of the conversation.
The specification evolves in dated versions. The one described in the documentation consulted, 2026-07-28, brings the protocol version and the client's capabilities to each request; a server announces its own by queryserver/discover. Two customer-side functions are deprecated, sampling (sampling) and roots (root). Optional extensions complete the core of the protocol: Tasks for long operations, MCP Apps to display an interactive interface in the conversation, and Skills over MCP to distribute agent instructions via MCP.
The protocol documentation quotes file servers, databases, GitHub, Slack or calendar. Some specific examples from the pages read:
@modelcontextprotocol/server-filesystem, used in the official tutorial for Claude Desktop: it reads, creates, moves and searches files, only in the folders you indicate to it;In the directory, companies describe their uses:ManoMano, an agent reads mobile application crashes by the official Firebase MCP server, limited to five read-only tools; Sorting agentWayflyerqueries internal tools by MCP;SNCF Connectpublished in ChatGPT an application based on the MCP Apps extension, with a server written in Kotlin. All linked sheets are listed below.
To build yours, the skill ofBuild a quality MCP serverguides the design of tools in Python (FastMCP) or TypeScript.
The principle is the same everywhere: either you declare the address of a remote server or the command that runs a local server, and then you authenticate if the service requests it. The screens and files change according to the software.
Open the application settings (from the Claude menu, not the account settings), Developer tab, Edit Config button. Fileclaude_desktop_config.jsonopen: declare the server under the keymcpServers, with the command to be launched (e.g.npxand the name of the package) and its arguments, then exit and restart Claude Desktop completely. The official tutorial prevents the server from running with the rights of your account: only open the folders you agree to see read and modified. Claude requests your agreement before each action.
In Settings, Connectors section, choose Add, then "Custom Connector", and paste the server address provided by its editor. After authentication, the connector settings allow to enable or disable each tool.
A remote server adds in one command, for exampleclaude mcp add --transport http notion . For a local server, the command that runs the server is placed after-- : claude mcp add --env AIRTABLE_API_KEY=VOTRE_CLE --transport stdio airtable -- npx -y airtable-mcp-server. The option --scopeselects the range:local(by default, you alone, in this project),project (a file .mcp.jsonshared with the team by the repository, which Claude Code has approved in interactive session before using it) oruser(all your projects). The command/mcpdisplays server status and launches OAuth authentication. SSE transport is depreciated to the benefit of HTTP.
Cursor lit a file .cursor/mcp.jsonin the project, or~/.cursor/mcp.jsonfor all your projects, with the same keymcpServers : command, args et envfor a local server,url et headersfor a remote server. Servers also install in one click from Cursor Marketplace. By default, Cursor requests your approval before using an MCP tool; in case of problem, the logs are in the Output panel, input "MCP Logs".
The "MCP: Add Server" command in the command palette launches a guided addition. VS Code-specific configuration format uses keyservers, where other tools usemcpServers. Documentation requires checking the server editor, the command or address and the accesses requested before installing or starting it.
Supabase publishes its server configuration for many clients, including Claude Code, Cursor and VS Code. In Claude Code:claude mcp add --scope project --transport http supabase "", puis /mcpto authenticate. In Cursor, the address states in.cursor/mcp.json. Added parameters to the address restrict access:project_reflimits the server to a project,read_only=trueonly allows reading,featureschooses the tool groups. The official skillDevelop with SupabaseAlso covers the CLI and the MCP server.
The specification clearly says: an MCP tool is about executing code, and the protocol cannot impose security rules itself. It is for applications and servers to apply them. Points that return to the pages read:
A MCP server gives the agent aaccess: data and actions in another system, via a programme that runs next to it. A skill (Agent Skills) gives him aknow-how: a folder that contains a SKILL.md file (a name, description, instructions) and sometimes scripts or resources. The agent remembers the name and description, and reads the rest only when the task is ready; Anthropic calls for progressive disclosure. Detailed explanation:What is a skill (SKILL.md)?
Both complement each other. In October 2025, Anthropic wrote to explore how skills can complement MCP servers, learning the agent more complex chains that use external tools; the MCP specification now lists an extension "Skills over MCP". In the library, the skillUse the Hugging Face Hub via its MCP serverserves only if this server is connected, andInstall the MCP Google Ads server guide son installation.
The two formats also share the same risk: a skill or a malicious server can cause the agent to leak data or act without anyone wanting to. Anthropic recommends installing skills only from trusted sources and auditing others, as for an MCP server.
Wayflyer, a Dublin fintech that finances growing brands, has built two agents: one prepares the diagnosis as soon as a ticket arrives, the other updates the procedures after its resolution.
At ManoMano, a chain of agents analyses the crashes of mobile applications every four hours, looks for whether the problem is already known and links each case to a ticket; the creation of tickets has been withdrawn at the AI.
SNCF Connect has released an application in ChatGPT to search for a train journey and is also developing its own conversational interface, to keep control of the customer relationship and data.
Connect a code agent to a design tool (here Paper, via MCP), then ask it to place an iOS mockup (a stock-trading app) directly on the canvas.
An agent connected to a canvas (here Askaboard, linked by MCP) draws, explains and poses exercises, to make understand a topic step by step.
List generated from library records whose name, summary or functions cite MCP; It extends with the new repositories.
claude-api
To be used as an up-to-date reference for building an application that calls the API Claude or an SDK Anthropic: model identifiers, parameters, streaming, tool call, MCP, agents, cache, token count and migration between models.
mcp-builder
Guides the creation of MCP servers that allow a model to act on external services with well-designed tools, in Python with FastMCP, or in Node and TypeScript with SDK MCP.
web-artifacts-builder
To create, in Claude.ai, ambitious, multi-component HTML artifacts with modern front technologies: React, TypeScript, Quick, Tailwind CSS and shadcn/ui, plus Parcel for clustering.
google-ads-api-mcp-setup
Guides the download, configuration and installation of the official and open Google Ads MCP server, to connect an assistant (Gemini, Claude Code, Cursor or equivalent) to a Google Ads account and query campaigns or reports in natural language.
google-ads-api-quickstart
From zero to a first query that lists the campaigns: token developer, ID and OAuth secret, refresh token, client account identifier, and if applicable login customer id.
agent-platform-architecture
Design and validation guide for a corporate Gemini architecture: multi-project hierarchy (host, central governance hub, execution radius, tool radius or MCP), Agent Gateway, connectivity models (all traffic or only private beaches), secure web proxy with policy routing or inter-cloud mTLS output hub, private certification authority and internal balancer, unified access policies (CEL) and semantic governance, third party identity provider (Entra ID, Ping, Okta) or SPIFFE agent identity management, Model Armor safeguards, inter-project registry synchronization, ADK agents on the runtime, and day operations 2.
agent-platform-rag-engine-management
This skill drives the RAG engine of Agent Platform: listing or inspecting corpus and files, recovering contexts, or generating content anchored in a corpus, via the SDK Python `vertexai` (Google GenAI), and not by raw REST calls or MCP tools, because the code is intended for external customers.
alloydb-basics
AlloyDB administration for PostgreSQL: clusters, instances, backups, and, if applicable, MCP AlloyDB tools to automate operations.
api-gateway-backend-security-and-rate-limiting
This skill configures client authentication (key API or JWT), gateway authentication to backend, and consumer project quotas, for OpenAPI 2.0 and 3.0.x configs, including an MCP public preversion gateway.
api-gateway-mcp-server
Configure Google Cloud API Gateway as remote MCP server, in public preversion, to expose API REST as agent tools on `POST By notifying an OpenAPI 3.0.x or 3.1.x specification with `x-google-api-management.mcp` and `x-google-mcp-tool`, the gateway translates JSON-RPC 2.0 requests (`initialize`, `notifications/initialized`, `tools/list`, `tools/call`) into HTTP calls to the backend, while maintaining authentication and quotas already in place.
api-gateway-openapi-spec-authoring
API Gateway accepts OpenAPI 2.0, 3.0.x and 3.1.x configs, but Google extensions do not have the same form from one version to another.
apigee-agent-gateway-governance
This skill sets up an Apigee proxy that an Agent Gateway Google Cloud calls in ext-proc (Service Extension), on the model traffic of an agent.
bigquery-migration-sql-translator
This skill relies on the BigQuery migration service (BQMS) and its MCP server to translate queries, DDL schematics and procedures stored from inherited dialects — Teradata, Oracle, Redshift, Snowflake, SparkSQL, HiveQL — to BigQuery's GoogleSQL.
cloud-monitoring-list-time-series-request
This skill translates a Cloud Monitoring metric descriptor into Call Settings ListTimeSeries ready for production: name, filter, interval, aggregation (alliner, reducer) and view.
cloud-monitoring-metric-selection
For a service or resource (Compute Engine, Spanner, BigQuery, Cloud Run, SQL Cloud, Pub/Sub, Cloud Storage, etc.), this skill directly searches metric descriptors via the MCP list metric descriptors tool, then filters them locally by keywords.
cloud-operator
Google Cloud operator and assistant via MCP autocloud server: practical questions, resources, Terraform, Kubernetes, Helm, deployments, incidents, costs.
datalineage-bigquery-asset-impact-analysis
When a BigQuery table or view is broken, out of date, or is being considered for modification, deletion or suspension of updates, this skill evaluates the downstream impact radius: tables, dashboards and dependent processes.
datalineage-summary
When you want to understand where and where data from a BigQuery or Cloud Storage asset goes, this skill runs through the lineage graph (Knowledge catalogue, MCP server) upstream and downstream, at the table or column level, and then tells it in a readable Markdown report, from left to right, without drowning the player under nodes and links.
detection-engineering-coverage-evaluation
End-to-end tooled feed for detection engineering in Google SecOps via MCP: extract intelligence from a ticket (without reading it in full in the answer), generate detection opportunities (TDO) on the cleaned text — without summarizing it before the call —, simulate the behavior of an attacker with synthetic UDM events, evaluate the coverage of each TDO (a separate parallel call, not an aggregate), and deploy rules that close holes.
gke-ai-troubleshooting-handle-disruption-gpu-tpu
When a GPU or TPU GKE node restarts, prepares for a Compute Engine host maintenance, or switches under a shutdown, this skill guides the diagnosis rather than creating a cluster.
gke-app-onboarding
For a first arrival on GKE, the skill continues to evaluate the application, containerization, image management (often to Artifact Registry, with analysis control), generation of manifests, and then deployment and verification.
gke-batch-hpc
For batch data pipelines, HPC simulations (fluid mechanics, molecular dynamics, financial models), parallel MPI or MapReduce calculation, ML drive or a CI build farm, this skill is directed to the Kubernetes Jobs, JobSet for multi-job workflows, and Kueue for queue (ClusterQueue, authorization of a namespace).
gke-cluster-creation
Rather than a single command, this skill chooses a model: Autopilot production (gold path, default for any new cluster), Autopilot dev/test, Regional standard for high availability or custom requirements, GPU inference (L4, ComputeClass) or AI Hypercompute.
gke-golden-path
The "gold path" is the recommended Autopilot configuration for a production cluster: sound default values, a checklist, and the requirement to document any deviation.
gke-inference
This skill deploys and adjusts IA inference loads on GKE, using GPU, TPU and model servers, relying on Google Inference Quickstart and LLM service practices.
gke-multitenancy
When several teams share a cluster, or when you isolate dev, staging and production in the same GKE, this skill sets a model: namespaces, RBAC at the team perimeter, resource quotas, LimitRanges, network insulation (including intraspace traffic and DNS), and allocation of costs per wording.
google-cloud-solution-guided-gke-ai-migration
Migration guide for an existing IA inference — Cloud Run, API Gemini, Gemini Enterprise Agent Platform, or VM custom — to a self-hosted inference on GKE, with gcloud and kubectl.
google-cloud-storage-basics
Cloud Storage (often called GCS) stores immutable objects in buckets, with strong overall consistency, several classes (Standard, Nearline, Coldline, Archive) and site types, plus IAM for fine access.
secops-cases
Google Security Operations SOAR folder life cycle: list, create, inspect, update or close a case, add comments and investigation notes, change priority or description, link or group alerts.
secops-detection-engineering
Detection engineering in Google SecOps: write, validate syntax, test (including a retrohunt on historical telemetry) and deploy YARA-L 2.0 rules, or evaluate coverage holes from intelligence (tickets, CVE, detection opportunities) with synthetic UDM events and long-term coverage analysis.
workload-manager-basics
Workload Manager evaluates loads in terms of general Google Cloud, SAP, SQL Server or custom organization rules.
retrieving-developer-knowledge
This skill questions the official developer Google documentation (Cloud, IA and Gemini, Android, Chrome, Web, Flutter, Go, Firebase, and other platforms) via the MCP Developer Knowledge server (answer query, search documents, get documents) or, in turn, the API REST developerknowledge.googleapis.com.
hf-mcp
To use the Hugging Face Hub through the MCP server tools: to search for models, data sets, Spaces and articles, consult a repository or documentation, start computing tasks, and use Spaces Gradio as the tools of the AIA.
cloudflare-email-service
The skill implements or troubleshoots the sending and routing of Cloudflare emails, as well as their delivery configuration.
web-perf
The skill audits, diagnoses or optimizes the loading and interaction of a site, Core Web Vitals and Lighthouse performance scores.
supabase
This skill applies to any Supabase task: base, Auth, Edge Functions, Realtime, Storage, vectors, cron, file, client libraries and SSR integrations (supabase-js, @supabase/ssr) in Next.js, React, SvelteKit, Astro or Remix, troubleshooting (session, JWT, cookies, getSession, getUser, getClaims, RLS), CLI or MCP server, schemas, migrations, audits, Postgres extensions, and debugging ( HTTP or Postgres errors, RLS, schema cache, timesouts, Logs Explorer / ClickHouse logs).
firebase-basics
This skill covers only the foundations of the Firebase LTC: installation, version control, connection (including --no-localhost), project creation and selection (firebase use), and downloading d-app configuration files (google-services.json, GoogleService-Info.plist).
eas-observe
Use everything that affects EAS Observe: add `expo-observe` to an Expo project (root HOC, interactivity markers, hook `useObserve`, Expo Router or React Navigation integrations, expo image and Reanimated, `Observe.logEvent` events, native errors and crashes, upload source maps, `sampleRate` configuration), and query metrics via the EAS CLI or MCP tools (summaries, routes, launches, errors).
neon
Presents Neon as a set of primitive backgrounds around Lakebase Postgres: authentication, object storage, functions, AI gateway and real-time queries, all instantaneous, plug-in and server-free.
neon-auth
To add authentication to a new application, with Neon Auth (better Auth managed): login, users and sessions that connect with the database, ID routing, registration, reset, OTP e-mail, magical links, organizations, OTP phone, OAuth, OAuth server for MCP or third party clients, access keys, MFA and trust domains.
neon-functions
Neon Functions deploys long-lasting HTTP Node.js functions on a Neon branch, with a public HTTPS URL, a calculation in the same region as the base, and automatic injection of DATABASE URL if the branch has Postgres.
docker-agent-config
Covers the agent.yaml (or .yml/.hcl) file of Docker Agent (open source cagent project, CLI docker agent): agents, models and suppliers, integrated tool sets or MCP, multi-agent teams with sub agents, and properties such as instruction, fallback or commands.
docker-agent-deploy
Covers the integration surface of Docker Agent once the configuration file is already written: expose the agent as a server (MCP, API HTTP, A2A, ACP or chat compatible OpenAI) with docker agent serve, distribute it via an OIC registry (docker agent share, for example to Docker Hub) and measure its quality with docker agent eval.
docker-docs
Used to answer questions about any Docker product from up-to-date official documentation, rather than agent's brief, including requests such as "how do I..." or "this still the right way".
docker-sandboxes-env
The `sbxenv.yaml` file (schemaVersion "1", still experimental) declaratively describes a Docker Sandboxes environment: agent, kits, mounts, arguments, variables, secrets, registers, links, MCP servers, ports and host commands `lifecycle:`.
claude-settings-audit
The skill analyses a repository to offer Claude Code (settings.json) permissions for read-only commands.
stripe-best-practices
Guides integration decisions Stripe: development and testing environment, choice of API, Connect, billing, tax, Treasury, integration options, deprecated migration of API, and security (keys, permissions, webhooks, OAuth).
browse-local-sentry
The skill connects an agent's browser to a local Sentry development server: either dev-it (pnpm run dev-ui) on an URL of the type org-slug.dev.getentry.net and a port, or the full devserver (devservices serve) in The deviation is critical: dev-it acts on the production sentry.io, with the user session, including a possible account staff or superuser.
sentry-debug-issue
To debug and correct a Sentry issue: find it (link, ID or search), extract the context (pile, breadcrumbs, trace, logs), possibly launch a cause analysis, apply the patch in the repository, and solve it via a `Fixes PROJECT-NAME-12A` commit or PR.
sentry-get-started
Guided point of entry to use Sentry via an agent: you first orient according to the configuration already in place, then, for a new project, you install Sentry from end to end with reasonable settings (provide a project, install the SDK, confirm the arrival of a real telemetry).
sred-project-organizer
The skill takes a list of projects and their documentation, then organizes them in the format expected for a SRED submission, with associated evidence.
sred-work-summary
The skill creates a year of work—typically the current calendar year, read through the date—by gathering GitHub pull requests, Notion documents and Linear tickets from a person, and then grouping these links by project into a Private Concept document, whose address it returns.
neon-postgres-agent-platforms
Used to design and operate a platform of agents or creators of multi-lease applications on Neon.
build-chatgpt-app
Allows you to build, scaffold, rework and diagnose a ChatGPT Apps SDK application, which combines an MCP server with a widget interface.
chatgpt-app-submission
Used to inspect the code of an MCP server of ChatGPT and to produce a `chatgpt-app-submission.json`: suggestions for app information, justifications for tool indices, test cases and negative test cases, and then report the review findings and warnings `outputSchema`.
ui-widget-developer
To build MCP servers for Copilot Chat, with widgets rendering (OpenAI Apps SDK or MCP Apps SDK, any language), including React and Fluent UI widgets and tools that return structuredContent.
copilot-sdk
Guides the creation of applications that rely on GitHub Copilot via the Copilot SDK, for programmatic integrations in Node.js/TypeScript, Python, Go or .NET: sessions, custom tools, streaming, hooks, MCP servers, BYOK providers, persistence, agents and skills, and deployment schemes.
mcp-builder
Guides the creation of MCP (Model Context Protocol) servers that allow models to interact with API or external services via well-designed tools.
agents-sdk
Guides building AI agents on Cloudflare Workers with SDK Agents: state-of-the-art agents, durable workflows, WebSocket real-time applications, scheduled tasks, MCP servers or chat applications.
ai-sdk
Provides expert guidance on the Vercel AI SDK for AI features: chat interfaces, text generation, structured output, tool call, agents, MCP integration, continuous stream, embeddings, reranking, image generation, or working with any model provider.
building-mcp-server-on-cloudflare
To build remote MCP servers on Cloudflare Workers, with tools, possible OAuth authentication, and deployment in production.
figma-design-to-code
Mandatory requirement, after the SKILL.md, before calling the MCP tool Figma get design context.
figma-use-figjam
The skill helps an agent to use the tool MCP use figima in a FigJam file (URL figma.com/board/...), in addition to figma-use for the basic context.
figma-use-motion
Provides the motion and animation context of the MCP tool use figma: manual key images, animation styles, smoothing and duration on the frieze.
vercel-connect
Vercel Connect guide to secure limited OAuth tokens for third-party services (Slack, GitHub, MCP, OAuth, Snowflake) on behalf of applications or users via OIDC Vercel.
assistant-mcp
Connects a code agent (Claude Code, Cursor, VS Code, OpenAI Codex) to Grafana Cloud via the MCP mcp-grafana server.
configuring-dbt-mcp-server
To produce the JSON configuration of the The server connects the wizard to the dbt CLI, Semantic Layer, the You start by choosing a local server (development stream, CLI access, typically via The tokens do not log, display, and ever s.
figma-to-content
Used to transform a Figma model into the AEM Edge Delivery Services content page (EDS / Franklin / Helix) in Document Authorizing (DA, da.live).
mongodb-mcp-setup
Guides the configuration of the MCP MongoDB server's essential options once installed.
figma-use
Mandatory prerequisites, according to the SKILL.md, before any call to the tool MCP use figma: Calling directly without this skill causes frequent failures and difficult to diagnose.
agent-framework-azure-ai-py
SDK Python Microsoft Agent Framework (agent-framework-azure-ai package) to build Azure AI Foundry agents.
airunway-aks-setup
The source text describes the implementation of AI Runway on an existing AKS cluster, from a naked cluster to a first model in service.
azure-ai
This skill guides the construction and troubleshooting of Azure AI Search, Speech, OpenAI and Document Intelligence applications.
azure-ai-voicelive-py
SDK Azure AI Voice Live for Python (azure-ai-voicelive), intended for applications that require real-time two-way audio communication with Azure AI: voice assistants, voice chatbots, live speech translation, voice-driven avatars, or any WebSocket audio stream with AI models.
azure-aigateway
Guides the configuration of API Management as a gateway for models, MCP tools and agents: semantic cache, token limit, content security, load distribution, model governance, MCP flow limitation, jailbreak detection, addition of Azure OpenAI backend or Foundry model, LLM policies, metrics and cost control, or conversion of API to MCP (OpenAPI import).
azure-kubernetes-automatic-readiness
The skill evaluates whether Kubernetes loads and the configuration of an existing cluster are compatible with AKS Automatic.
azure-kusto
This skill is used for the query of Azure Data Explorer (Kusto/ADX) in KQL: writing and executing KQL queries for log analytics, telemetry, time series, IoT telemetry, or detection of anomalies.
azure-quotas
This skill is used for consultation and demand for Azure quotas, as well as for routine use, to plan a deployment or choose a region: check service limits, use, and request an increase — this demand is free: only the resources actually used are paid, not the quota allocation.
deploy-model
Unified deployment skills of Azure OpenAI models, which are directed according to the intent: preset rapi deployment, fully customized deployment (version, SKU, capability, RAI policy) or capacity discovery between regions and projects.
airtable-cli
To list databases, read and write records, manage tables and fields, filter and search for Airtable data, via the `airtable-mcp` CLI.
airtable-filters
Used to find, filter or circumscribe Airtable records according to field values, even when the word "filter" is not pronounced.
clean-up-dropbox-content
Use to clean Dropbox content by finding obsolete, duplicate, temporary or unwanted files, and delete only after an explicit review.
connect-chatcut-desktop
Serves to install, open, connect or repair ChatCut Desktop when MCP chatcut desktop tools are missing or unavailable.
vercel-api
This skill helps to question projects, deployments, environment variables, domains, newspapers or Vercel documentation, either via the connected application (MCP, HTTP Streamable Transport to ` or via the API REST, the SDK `@vercel/sdk`, or the `vercel api` command that appeared in January 2026.
web-perf
Analyze the performance of a page with the MCP Chrome DevTools.
implementing-api-surveys
Implements a survey of the Use when you have asked to run a survey, build a headless survey, build a custom survey IU, make a survey in my application, or run a Covers posthog-js event contract, server-side capture, and post-implementation verification by To design or create the survey definition, use creative surveys instead.
migrating-from-fivetran
Move an user's Fivetran connectors to PostHog data warehouse sources.
understanding-billing-usage
Explains the use of PostHog billing and the expenses of the client's visible billing tools clarify.
diagnosing-ci-and-merge-bottlenecks
Diagnosis CI and the demand for a health pipeline for a GitHub repo with engineering analysis tools Use when you are asked if CI becomes faster or slower, what kind of GitHub Actions work flow is slow or failing, how long do PR take from opening to merging, which open PR failed or waiting for a CI, where a draw request is blocked, what cost a failed test, which team has the most flakiest tests, or this cost of CI for a PR or workflow.Triggers on "engineering analysis", "ICR becomes slower", "low workflow", "ICR is low", "merger time", "cycle time", " Renewal flow", "fail controls", "where is stuck PR<n>"The long pole of IC", "The cost of IC".
exploring-mcp-sessions
Investigating the individual sessions of PostHog — the sequence of tool calls from a single agent does at one time, what he was trying to do, and where he was wrong.
migrating-from-customer-io
Move Customer.io automations to PostHog as a workflow.
building-a-dashboard
Build a new dashboard, or update an existing board, from a set of information — the same work asTRA in the application done with its high-voltage dashboard tool, but above Use when an user requests to create a dashboard, put multiple metrics/graphics on a page, assemble a dashboard for a topic (product analysis, retention, revenue, activation, etc.), or add/delete/replace information on a dashboard he already has.
exploring-mcp-tool-quality
Study the quality of PostHog tool calls – error rate,volving Latency, range, and which tools are failing or slow.
checking-member-access
Explains what a member or role can do in a PostHog project, using the access control tools Use when the user asks what someone can see or change, who can change dashboards or flags, why a member can or can't open a dashboard, notebook or table, what a role grants, what properties are hidden from someone, or how default project access is defined.
exploring-mcp-tool-original-user-motive
Build a starting taxonomy for a tool – what users were trying to accomplish before reaching the tool – and publish it as a PostHog notebook.
working-with-task-comments
Read and use the comments attached to the current PostHog task, its artifacts, and any canvas through the exec PostHog Use when the user mentions task comments, objects or canvases, annotations, selected text comments, answers, unsolved comments, or ask an agent to inspect or act on comments left in PostHog Desktop.
exploring-llm-traces
Debug and inspect the traces of LLM/AI agents using PostHog's tools.
creating-surveys
Creates and initiates post-hog surveys through the K0, including NPS/CSAT popovers, hosted feedback forms, and headless surveys.
downloading-batch-export-files
Export PostHog events, people, sessions or results of a HogQL query on request and download the resulting Shortly.
designing-email-templates
Author, backup, and edit email templates in PostHog workflow library — Compose JSON email design with Liquid customization and create and return edit templates on Use when asked to design, build, update or fix an email template for workflows, programs or campaigns.
exploring-mcp-intent-clusters
Exploring PostHog Intent Clusters - Agent Objectives grouped by semantic similarity, with tool distribution and error rates of each cluster, plus tool centered pivot (intention capture rate, discovery rate compared to the advertised catalogue, description adjustment, tool overlaps).
exploring-apm-traces
Survey of the performance of distributed applications using PostHog APM (OpenTelemetry spread) data via Use when requesting service traces, slow HTTP/basendical scales, error scales, trends or peaks of error rates, latency distributions, trace identifiers or range attributes — not trace of AI or product logs.
exploring-mcp-tool-usage
Starting point to explore how the tools of a PostHog server are used, K0-K — directs a general question to the typed tool that answers it.
managing-streamlit-apps
Create, deploy, and operate Streamlit applications in PostHog via streamlit-apps tools - create an application, set its source, read and correct its code, start and stop its sandbox, survey status, list versions, delete and share the application with humans via its PostHog URL.
querying-canvas-data
Get PostHog data into a canvas correctly: the host-injected Use every time a canvas displays metrics, graphs, tables, or PostHog data, or GitHub or server data, or send analytical events.
signals-scout-mcp-tool-calls
Signal Detector for PostHog tool calls Timepieces for tools that need improvement — wide range failure rates, back to action, slow or contextual responses — grouped by proprietary product category, each with a fixed suggestion.
modeling-warehouse-foundations
Shared databases to build reusable data models in PostHog, on either of the two piles: PostHog-native data-warehouse views / materialized views (HogQL, via the tool(s) Read before writing a specific business model — covers the decision PostHog-vs-dbt, the view-create → view-materialize → sync frequency workflow and the aliasing rule of the HogQL column, the skeleton of the dbt project and the image "native dbt integration", the warehouse joins and the star-schema dimensions, the currency conversion with convertCurrency(), and the verification/recording of the models in the data catalog for reuse.Companion to modelling, modelling, conversion, modelling, activation, modelling, product use and modelling, dimension.
investigating-logs
Investigating newspapers in a PostHog project: checking that a service or deployment is healthy, explaining a peak of error, sorting an incident, or understanding what a newspaper stream says.
managing-path-cleaning-rules
Inspects URL paths and offers, tests, controls and applies path cleaning rules at the project level so that dynamic segments (digital ID, UUIDs, slugs, dates) collapse into readable aliases.
feature-flags-web
PostHog Options for Web Applications (JavaScript)
feature-flags-java
PostHog Options for Java Applications
feature-flags-python
PostHog Options for Python Applications
feature-flags-android
PostHog Options for Applications
feature-flags-ios
PostHog Options for Applications
feature-flags-rust
PostHog options for Rust applications
feature-flags-go
PostHog Options for Go Applications
feature-flags-react
PostHog Options for React Applications
feature-flags-react-native
PostHog Functional Flags for Native Applications React
feature-flags-api
PostHog Options for Applications
feature-flags-dotnet
PostHog options for .NET applications
feature-flags-php
PostHog Options for PHP Applications
feature-flags-flutter
PostHog Options for Applications
feature-flags-ruby
PostHog Options for Ruby Applications
feature-flags-nextjs
PostHog Options for Next.js Applications
feature-flags-elixir
PostHog Options for Elixir Applications
feature-flags-nodejs
PostHog options for Node.js applications
A program that provides tools, data and instructions to an AI application according to the MCP (Model Context Protocol) standard protocol. It can run on your computer or on the Internet.
Yes: "MCP server" is the translation of "MCP server". MCP is the acronym for Model Context Protocol, the open source protocol published by Anthropic on November 25, 2024.
Not to connect one that exists: in Claude, a custom connector is created by sticking the server address, and Cursor offers an one-click installation. A local server asks to edit a JSON file and often install Node.js, as in the official tutorial of the Filesystem server. To build one, you have to program with an SDK; the skillBuild a quality MCP serveraccompany this step.
In Claude on the web or computer: Settings, Connectors, Add, "Custom Connector", then the server address. In Claude Desktop, a local server appears in the claude desktop config.json file (Settings, Developer, Edit Config). In Claude Code: command claude mcp add, then /mcp to authenticate.
Its hosted address is https://mcp.supabase.com/mcp. In Claude Code: claude mcp add --scope project --transport http supabase "https://mcp.supabase.com/mcp", then /mcp to authenticate. Supabase recommends limiting the server to a project (project ref), enabling read-only (read only=true) and not entrusting it to your end clients.
It depends on the server and how to connect it. The specification specifies that the protocol cannot impose security itself: a tool is to execute code. Install reliable source servers, keep each tool call approved, give as little rights as possible, and keep keys out of the conversation.
The MCP server gives access to a system (data, actions); the skill brings instructions and scripts that the agent loads when the task is ready. Both combine: some skills assume that an MCP server is connected. SeeWhat's a skill?
All read on 10 November 2026. The guide text is original writing by iadujour.fr, not a translation.