Best MCP servers 2026: 15 picks with security caveats

Gives an agent full access to GitHub's API surface: reading and writing issues, pull requests, discussions, code search, and branch management. Backed by GitHub's own identity and permissions model, which sets it apart from community wrappers.

When to use: Any agentic coding workflow where the agent needs to open PRs, file issues, review diffs, or search across repositories without you doing it manually.

Controls a real browser via Playwright: navigate pages, click, fill forms, take screenshots, and extract structured content using accessibility snapshots. Does not require a vision model; the agent works from the accessibility tree.

When to use: QA automation, scraping sites that block curl, filling forms as part of an agentic workflow, or visual regression testing. The accessibility-snapshot approach is faster and cheaper than screenshot-based agents.

Sandboxed read and write access to local files and directories. Configurable path allowlist prevents the agent from touching files outside the declared roots. Part of Anthropic's official reference server set.

When to use: Any agent that needs to read, write, or organize files on disk — code generation, document processing, config management. The path allowlist is the primary guardrail; configure it before anything else.

Knowledge-graph-based persistent memory stored locally as JSON. The agent can store entities, relationships, and observations across sessions. No external API calls; data stays on the machine running the server.

When to use: Long-running agentic workflows where context needs to survive across sessions — project tracking, relationship mapping, accumulating research findings. The local-storage model makes it a reasonable starting point before committing to a vector database.

Connects an agent to a Supabase project for natural-language database queries, table inspection, and schema reads. Supports both the anon key (limited) and the service-role key (full access).

When to use: Data analysis, debugging production data issues, scaffolding queries during development. Pair with read-only credentials for any production database.

Exposes Stripe's API as MCP tools: check subscription status, create invoices, manage customers, query revenue. Stripe also hosts a remote MCP endpoint at mcp.stripe.com if you prefer not to run it locally.

When to use: Revenue dashboards, automated billing workflows, customer support agents that need to check subscription state without manual API calls. The remote endpoint removes the local dependency but introduces a network trust boundary.

Real-time web search, URL extraction, site mapping, and crawl tools built for AI agents. Tavily returns structured, LLM-ready results rather than raw HTML. Also available as a hosted remote endpoint.

When to use: Research agents, fact-checking pipelines, news monitoring, competitive analysis. More reliable than Puppeteer for basic search because it handles rate limits and anti-bot measures at the API layer.

Web and local business search via the Brave Search API. Brave does not profile or track individual queries, which matters if your agent is processing sensitive research topics. Returns web results and local business data.

When to use: General-purpose search in workflows where query privacy matters, or as a lower-cost alternative to Tavily when you need basic web results without crawling depth. Requires a Brave Search API key.

Natural-language query execution against a PostgreSQL database. Defaults to read-only mode, which prevents accidental mutations. Schema introspection lets the agent understand table structure before writing queries.

When to use: Ad-hoc data analysis, debugging slow queries, building reports without writing raw SQL. Keep it read-only unless you have a specific, well-scoped write task and a rollback plan.

Gives agents access to AWS documentation, service metadata, billing data, and CDK scaffolding. Covers a wide surface: S3, Lambda, CloudWatch, Cost Explorer, and more. Actively maintained by AWS Labs.

When to use: Infrastructure agents, cost analysis, documentation lookup during IaC development. Pairs well with Terraform MCP for a full infrastructure workflow.

Reads Sentry error events, traces, and performance telemetry through an MCP interface. An agent can triage issues, look up stack traces, and correlate errors with deploys without leaving the coding environment.

When to use: Debugging agents that need production error context, or any workflow where you want the agent to automatically check Sentry before proposing a fix. Substantially cuts time-to-diagnosis on production incidents.

Gives agents structured access to the Terraform registry: provider documentation, module schemas, resource definitions. The agent can look up resource arguments and generate correct HCL without hallucinating attribute names.

When to use: IaC generation, module discovery, debugging Terraform plan errors. This server is read-only against the public registry — it does not run terraform apply. Pair with a local Terraform execution step that you control.

Connects agents to MongoDB and Atlas clusters. Supports structured queries, schema introspection, and aggregation pipelines. Built-in auth and access control support, which puts it above most community MongoDB wrappers.

When to use: Document-store analytics, debugging MongoDB query performance, natural-language data exploration on Atlas clusters. The official auth integration makes it safer to use against real data than a community wrapper.

Fetches web content and converts it to Markdown for efficient processing by the model. Handles basic HTML-to-text extraction, which covers the majority of documentation and article reading use cases without spinning up a full browser.

When to use: Documentation lookup, reading articles, pulling changelog pages during research. Use Playwright MCP when JavaScript rendering is required; Fetch MCP when it is not.

Read, create, and update Notion pages, databases, and blocks through the Notion API. Useful for agents that need to write findings to a shared workspace, update project databases, or pull structured data from Notion tables.

When to use: Project management agents, documentation generators, any workflow where the output needs to land in a shared Notion workspace. Rate limit is 3 requests/second; batch writes accordingly.