The best MCP servers for developers and DevOps in 2026

Updated November 15, 2026By HostCrust12 tools compared17 min read

Start with three: GitHub's official MCP server for repositories, issues and pull requests (it has a real read-only mode), Context7 for current library docs (it reads documentation and can't change anything), and Microsoft's Playwright MCP so the agent can test its own work in a browser.

Then add the official server for your stack: Sentry, Grafana, Kubernetes, AWS, Cloudflare or your database. Turn on each one's read-only switch and give it its own narrowly scoped credential. The token's permissions, not the server's tool list, are what really limit an agent.

Our top picks

Compare at a glance

ToolBest forPricingOpen sourceDeploymentTested by us
GitHub MCP Serverrepositories, issues, pull requests and ActionsFree (open source); it acts with your GitHub account's permissionsYesRemote (hosted by GitHub) or self-hosted local serverDocs only
Context7current library docs inside the agentFree plan (1,000 API calls a month, public repositories); Pro $10 per seat per month; Enterprise customNoSaaS (remote MCP endpoint), or a local MCP server or CLI that calls itDocs only
Playwright MCPletting the agent test its work in a real browserFreeYesSelf-hosted (npm package or Docker)Docs only
Sentry MCP Servergoing from a production error to a fixMCP server free; needs a Sentry plan: Developer free (1 user), Team $26 a month billed annually ($29 monthly), Business $80 a month billed annually ($89 monthly)NoSaaS (remote at mcp.sentry.dev) or self-hosted stdio serverDocs only
Grafana MCP server (mcp-grafana)metrics, logs, dashboards and alertsFree (open source)YesSelf-hosted (binary, uvx or Docker)Docs only
Kubernetes MCP ServerKubernetes and OpenShift clustersFreeYesSelf-hosted (release binary, npm or PyPI package)Docs only
AWS MCP ServerAWS accounts, with IAM and CloudTrail behind itNo additional charge; you pay for the AWS resources your agents useNoSaaS (managed by AWS)Docs only
Cloudflare API MCP ServerCloudflare DNS, Workers and account settingsFree (open source); actions use your Cloudflare account and planYesSaaS (remote at mcp.cloudflare.com)Docs only
MCP Toolbox for DatabasesPostgreSQL and other databasesFreeYesSelf-hosted (binary or container)Docs only
Filesystem MCP server (reference)file access for chat apps without their own file toolsFreeYesSelf-hosted (npm package or Docker)Yes
Git MCP server (reference)Git for agents that can't run git themselvesFreeYesSelf-hosted (PyPI package via uvx or pip, or Docker)Yes
ssh-mcp-guard Oursapproved changes on Linux servers over SSHFreeYesSelf-hosted MCP server plus a server-side gateYes

The full ranking

GitHub MCP Server

Best for: repositories, issues, pull requests and Actions

GitHub's own server is the one most developers should add first. It's actively maintained (2.1.0 on 3 Nov 2026, MIT) and runs either as a remote server GitHub hosts or as a local binary or Docker image, signing in with OAuth or a personal access token. It also has the controls a busy server needs. --toolsets loads only what you use (the default is context, repos, issues, pull_requests and users). --read-only drops every write tool, even ones you list by name. A lockdown mode hides content in public repositories written by people without push access, which cuts prompt-injection risk. GitHub says lockdown isn't an authorization boundary, so scope the token as well.

Pricing
Free (open source); it acts with your GitHub account's permissions (checked November 15, 2026, source)
License
Open source (MIT)
Deployment
Remote (hosted by GitHub) or self-hosted local server
Runs on
Any MCP client; remote endpoint or a local binary or Docker image on macOS, Linux and Windows

Pros

  • Official and very active
  • Read-only mode the server enforces, plus toolsets to cut what loads
  • Remote with OAuth, or local for GitHub Enterprise Server

Cons

  • Acts with everything the token or OAuth grant allows
  • Issue and pull request text from strangers can carry prompt injection
  • GitHub Enterprise Server can't use the remote server

Not tested by us yet: based on the vendor's documentation.

Context7

Best for: current library docs inside the agent

Context7 fixes a daily annoyance: agents writing code against an API version that no longer exists. Its two MCP tools, resolve-library-id and query-docs, fetch version-specific docs and examples from Upstash's index. It's low risk by design, because it reads documentation and touches nothing in your environment. Its own README adds two caveats. The documentation is community-contributed and not guaranteed accurate or secure. The backend (API, parser and crawler) is private, so only the thin MCP client is open source. The free plan allows 1,000 API calls a month; Pro is $10 per seat. Upstash also offers a CLI-plus-skill mode that skips MCP.

Pricing
Free plan (1,000 API calls a month, public repositories); Pro $10 per seat per month; Enterprise custom (checked November 15, 2026, source)
License
MCP server open source (MIT); the documentation API behind it is proprietary
Deployment
SaaS (remote MCP endpoint), or a local MCP server or CLI that calls it
Runs on
Any MCP client; remote endpoint at mcp.context7.com

Pros

  • Can't change anything in your environment
  • Cuts made-up and outdated API calls
  • Free plan, remote endpoint, nothing to run

Cons

  • The documentation backend is proprietary
  • Community-contributed docs can be wrong or carry injected instructions
  • Free plan is capped at 1,000 calls a month

Not tested by us yet: based on the vendor's documentation.

Playwright MCP

Best for: letting the agent test its work in a real browser

Microsoft's Playwright MCP lets an agent open pages, click, fill in forms and read the result through the page's accessibility tree, with no screenshots or vision model needed. That makes it the usual way for a coding agent to check its own front-end changes. It's also the server here that needs the most care. Its README says it's “not a security boundary”. --allowed-origins doesn't cover redirects, and by default it keeps a persistent browser profile, so logins survive between sessions. Use --isolated, a separate profile and test accounts. Microsoft now suggests coding agents try its Playwright CLI with skills, which uses fewer tokens.

Pricing
Free (checked November 15, 2026, source)
License
Open source (Apache-2.0)
Deployment
Self-hosted (npm package or Docker)
Runs on
macOS, Linux and Windows with Node.js 18+

Pros

  • Official and very active (0.0.85 on 9 Nov 2026)
  • Works from the accessibility tree, so no vision model
  • File access limited to the workspace roots by default

Cons

  • Not a security boundary, by Microsoft's own account
  • Persistent browser profile unless you pass --isolated
  • Every page it reads is untrusted input

Not tested by us yet: based on the vendor's documentation.

Sentry MCP Server

Best for: going from a production error to a fix

Sentry's hosted server at mcp.sentry.dev gives a coding agent your issues, stack traces, traces and releases, so it can go from an error to a fix without you pasting logs. You sign in with OAuth and can narrow the tools with skills (?skills=inspect,triage on the URL or MCP_SKILLS locally). A stdio mode serves self-hosted Sentry. Two things to watch. When you run it yourself, its AI search tools need an LLM provider key. And the token scopes its README lists include writes (project:write, team:write, event:write), so prefer OAuth or a trimmed token. The code is source-available under FSL-1.1, which turns into Apache-2.0 two years after each release.

Pricing
MCP server free; needs a Sentry plan: Developer free (1 user), Team $26 a month billed annually ($29 monthly), Business $80 a month billed annually ($89 monthly) (checked November 15, 2026, source)
License
Source available (FSL-1.1-Apache-2.0)
Deployment
SaaS (remote at mcp.sentry.dev) or self-hosted stdio server
Runs on
Any MCP client; Sentry SaaS or self-hosted Sentry

Pros

  • Official and active (0.49.0 on 7 Nov 2026)
  • Remote with OAuth, so no token in your config
  • Skills narrow which tools load

Cons

  • Source-available rather than open source
  • Suggested token scopes include write access
  • Self-run AI search tools need an LLM provider key

Not tested by us yet: based on the vendor's documentation.

Grafana MCP server (mcp-grafana)

Best for: metrics, logs, dashboards and alerts

Grafana Labs' server (2.3.0 on 10 Nov 2026, Apache-2.0) lets an agent search dashboards, run data source queries such as Prometheus and Loki, and read alerts and incidents. It's our observability pick because its controls are good. Every tool category has a --disable-… switch, and --disable-write removes the create and update tools. With --disable-api set, --prometheus-enforced-matchers and --loki-enforced-matchers add a label filter to every native Prometheus and Loki query, which limits what it can read; the README warns that rendering and Assistant tools can still bypass the filter unless you turn them off too. It signs in with a service account token, and the README lists the RBAC permissions each tool needs. Grant those, not the Editor role it offers as a shortcut.

Pricing
Free (open source) (checked November 15, 2026, source)
License
Open source (Apache-2.0)
Deployment
Self-hosted (binary, uvx or Docker)
Runs on
macOS, Linux and Windows; works with self-hosted Grafana or Grafana Cloud

Pros

  • Per-category switches and --disable-write
  • Enforced label matchers limit which series and logs it can read
  • Works with self-hosted Grafana and Grafana Cloud

Cons

  • Many tools load by default, which costs context
  • The suggested Editor-role shortcut grants writes
  • Query access can expose sensitive log lines to the model

Not tested by us yet: based on the vendor's documentation.

Kubernetes MCP Server

Best for: Kubernetes and OpenShift clusters

This server (0.1.3 on 13 Nov 2026, Apache-2.0) talks to the Kubernetes API directly rather than wrapping kubectl. It covers pods, logs, events, any resource and Helm, across every context in your kubeconfig. Its TOML config has read_only = true, toolsets and denied_resources (for example, Secrets). The README calls these best-effort guardrails: the control to rely on is Kubernetes RBAC for the identity it uses. So give it a dedicated kubeconfig with view-only roles, not your cluster-admin context. Settings moved from command-line flags to TOML, so older setup guides are out of date.

Pricing
Free (checked November 15, 2026, source)
License
Open source (Apache-2.0)
Deployment
Self-hosted (release binary, npm or PyPI package)
Runs on
macOS, Linux and Windows; manages Kubernetes and OpenShift clusters through kubeconfig or in-cluster config

Pros

  • read_only setting and a denied_resources list
  • Many clusters, plus OpenShift
  • Calls the API directly instead of shelling out to kubectl

Cons

  • Its guardrails are best-effort; RBAC is the real limit
  • Your default kubeconfig is often an admin context
  • Pod exec and Helm install tools can change the cluster

Not tested by us yet: based on the vendor's documentation.

AWS MCP Server

Best for: AWS accounts, with IAM and CloudTrail behind it

AWS's managed server became generally available on 6 May 2026 as part of the Agent Toolkit for AWS, which AWS calls the successor to its older awslabs MCP servers. Agents get documentation search and AWS API access through aws___run_script, which runs Python in a sandbox AWS hosts, with no access to your local files or shell. The safety controls sit on AWS's side. IAM condition keys let you write policies that apply only to agent actions, for example read-only through the MCP server even if your role can write. Every request is logged in CloudTrail. There's no extra charge, and the service itself runs in US East (N. Virginia) and Europe (Frankfurt).

Pricing
No additional charge; you pay for the AWS resources your agents use (checked November 15, 2026, source)
License
Proprietary managed service (the local mcp-proxy-for-aws is Apache-2.0)
Deployment
SaaS (managed by AWS)
Runs on
Any MCP client through mcp-proxy-for-aws; US East (N. Virginia) and Europe (Frankfurt)

Pros

  • Limits enforced by IAM, not by the client
  • CloudTrail logging and CloudWatch metrics for agent activity
  • Scripts run in an AWS sandbox, away from your machine

Cons

  • AWS only, and a proprietary service
  • A broad IAM role still means broad power
  • AWS now steers users from the local awslabs servers to this one

Not tested by us yet: based on the vendor's documentation.

Cloudflare API MCP Server

Best for: Cloudflare DNS, Workers and account settings

Cloudflare's remote server covers its whole API, about 2,500 endpoints, with four tools: search and execute do the API work, plus docs and whoami. It uses what Cloudflare calls code mode: the agent writes code that searches the API spec and calls endpoints, and only the results come back. Cloudflare puts that at roughly 1,100 tokens of tool definitions, against about 244,000 for one tool per endpoint with minimal schemas (about 1.17 million with full schemas). You connect with OAuth and choose permissions, or use an API token. If a call needs a scope you didn't grant, the server asks the client to re-authorize. That makes the OAuth screen your main safety control: grant read scopes unless you mean the agent to change DNS or Workers. Several of Cloudflare's older product-specific servers are marked deprecated in its older repository, which asks users to move to this one.

Pricing
Free (open source); actions use your Cloudflare account and plan (checked November 15, 2026, source)
License
Open source (Apache-2.0)
Deployment
SaaS (remote at mcp.cloudflare.com)
Runs on
Any MCP client with remote-server support; a Cloudflare account

Pros

  • Whole API in a small context
  • Permissions chosen on Cloudflare's OAuth screen
  • Hosted by Cloudflare, nothing to install

Cons

  • Agent-written code runs against your account
  • A broad grant reaches DNS and security settings
  • Remote only

Not tested by us yet: based on the vendor's documentation.

MCP Toolbox for Databases

Best for: PostgreSQL and other databases

Google's MCP Toolbox (1.17.0 on 13 Nov 2026, Apache-2.0) is the most actively maintained database server we found. Start it with --prebuilt=postgres (or MySQL, SQL Server, AlloyDB, BigQuery and others) for ready-made tools: list tables, indexes, locks and long-running transactions, or run SQL. Or define your own tools in tools.yaml so the agent can only call the queries you wrote. Its docs are blunt: tool annotations and client confirmations are not a database security boundary. Connect it as a dedicated role with only the privileges the agent should use, ideally SELECT-only, never an owner or admin account.

Pricing
Free (checked November 15, 2026, source)
License
Open source (Apache-2.0)
Deployment
Self-hosted (binary or container)
Runs on
macOS, Linux and Windows; PostgreSQL, MySQL, SQL Server, AlloyDB, Cloud SQL, BigQuery and other databases

Pros

  • One server for PostgreSQL and many other databases
  • Custom tools can limit the agent to fixed queries
  • Very active, with clear security guidance

Cons

  • The SQL tool runs anything the database role allows
  • Safety depends entirely on the role you connect with
  • Results can expose personal data to the model

Not tested by us yet: based on the vendor's documentation.

Filesystem MCP server (reference)

Best for: file access for chat apps without their own file tools

The MCP project's Filesystem server (2.3.0 on 11 Nov 2026) reads and writes files only inside the directories you pass on the command line. Its newer switches are worth using. --read-only serves only the read tools, and --read-only-dir protects some directories while others stay writable. In our test both worked as documented, and a symlink pointing outside the allowed directory was refused. There are two cautions. The repository calls these reference implementations, not production-ready, and this server has had 8 published security advisories, mostly ways around its directory checks. And coding agents such as Claude Code already have file tools, so it's mainly useful for chat apps.

Pricing
Free (checked November 15, 2026, source)
License
Open source (Apache-2.0 for new code; older contributions MIT)
Deployment
Self-hosted (npm package or Docker)
Runs on
macOS, Linux and Windows with Node.js

Pros

  • Read-only modes that worked in our test
  • Symlinks out of the allowed directories refused in our test
  • --confine-roots stops a client from widening access

Cons

  • A reference implementation, not production-ready
  • 8 published security advisories so far
  • Write tools load unless you ask for read-only

Tested by us: Version 2.3.0 from npm on macOS 14.5, 14 Nov 2026 (UTC): with --read-only it listed 10 tools (no write_file, edit_file or move_file) and a write_file call failed as an unknown tool; with --read-only-dir it wrote to the writable directory but refused the protected one; a read through a symlink pointing outside was refused.

Git MCP server (reference)

Best for: Git for agents that can't run git themselves

The reference Git server (2026.11.11 on PyPI) has 14 tools. Eight are marked read-only, such as status, diff, log and show. The rest are commit, add, reset, create branch, checkout and merge. It has no push, clone or read-only switch. Always start it with --repository: without it, its README warns, it reads and writes any repository the process can reach. In our test it refused a repository outside that path and rejected option-style arguments. Its history argues for caution: 7 published advisories, including a High-severity one in which a repository's own config could run commands during read-only tools. If your agent can run git itself, you may not need it.

Pricing
Free (checked November 15, 2026, source)
License
Open source (MIT, moving to Apache-2.0 with the rest of the repository)
Deployment
Self-hosted (PyPI package via uvx or pip, or Docker)
Runs on
macOS, Linux and Windows with Python

Pros

  • --repository confinement held in our test
  • No network tools (no push, fetch or clone)
  • Won't run programs named in a repository's config or hooks, per its README

Cons

  • No read-only switch
  • 7 published security advisories so far
  • Redundant for agents that already have a shell

Tested by us: Version 2026.11.11 from PyPI on macOS 14.5, 14 Nov 2026 (UTC), started with --repository: git_log worked inside the repository, git_status on a second repository was refused as outside the allowed repository, and git_checkout and git_diff calls with option-style arguments (--orphan=x, --output=/tmp/x) were rejected.

ssh-mcp-guard Ours

Best for: approved changes on Linux servers over SSH

We make this one, so weigh that. It fills the SSH slot. Read-only diagnostics run straight away. Every change (a restart, a patch or a reboot) needs an approval signed with your SSH key, and the server checks it through an sshd forced command, so the agent can't get around it with plain ssh and the same key. That server-side enforcement is the one criterion it wins on. It ranks last on maturity: version 0.9.0 with no outside security review, setup on every server, OpenSSH 8.1+ for approvals (AlmaLinux and Rocky 8 are read-only), and a fixed catalog of 11 actions with no file transfer. Our SSH MCP servers ranking covers the alternatives.

Pricing
Free (checked November 15, 2026, source)
License
Open source (Apache-2.0)
Deployment
Self-hosted MCP server plus a server-side gate
Runs on
Linux servers with OpenSSH 8.1+; the MCP server runs on macOS or Linux

Pros

  • Policy enforced on the server by sshd
  • Each change needs a signed, single-use approval
  • Hash-chained receipts you can check offline

Cons

  • Setup on every server
  • A fixed catalog of 11 actions and no file transfer
  • Version 0.9.0, not yet independently reviewed

Tested by us: Ours, and run end to end on our Debian 12 and AlmaLinux 9.8 test servers (14 Nov 2026, UTC): signed restart and patch, refused critical and never-started units, refused replays and a missing nonce store. No independent review of the code yet.

How we picked

We started from the official MCP Registry, the MCP project's own reference servers and the servers that vendors such as GitHub, Microsoft, Sentry, Grafana Labs, AWS, Cloudflare and Google publish for their own systems. On 15 Nov 2026 we checked each one's repository for its license, latest release and latest commit, and read its README and security advisories. We ranked on five things, in this order:

  • Maintained: a release or commit within the last two months. Postgres MCP Pro, for example, has had no release since May 2025, so it's listed under other servers rather than ranked.
  • Official: built by the vendor whose system it touches, or by the MCP project itself.
  • Useful every week to developers and DevOps people working in Claude Code, Codex or Cursor.
  • Least-privilege controls: a read-only mode, ways to load fewer tools, scoped credentials, and above all limits enforced by the system the server talks to rather than by the server alone.
  • Track record: published advisories, and whether the vendor writes down the server's limits.

We don't give scores. Where two servers are close, the order reflects how many readers will use them. Pick the vendor servers that match your stack rather than reading the middle of the list as a strict ranking. We ran the Filesystem and Git reference servers ourselves (see what we tested); the rest are reviewed from their own documentation.

Security notes: what each MCP server can touch

An MCP server is code that acts for the model with your credentials. Before adding one, check three things: what it can reach, whether it can write, and where the limit is enforced. A limit inside the MCP server only holds while the agent goes through that server. A limit in the GitHub token, the IAM policy, the Kubernetes role or the database role holds even if the agent finds another way in.

ServerWhat it can reachWrites by default?Built-in limitLimit that really holds
GitHubRepos, issues, pull requests, Actions the grant allowsYes--read-only, toolsets, lockdown modeToken or OAuth scopes
Context7Upstash's documentation indexNo (reads docs only)None neededNothing of yours is exposed
PlaywrightAny website, your browser profile, workspace filesYes (clicks, forms)--isolated, origin lists (not a boundary)Test accounts and a separate profile
SentryYour Sentry organization's issues and eventsSome toolsSkillsOAuth grant or token scopes
GrafanaDashboards, data sources, alerts, incidentsYes--disable-write, enforced matchersService account RBAC
KubernetesAny resource, pod exec, HelmYesread_only, denied_resources (best-effort)Kubernetes RBAC
AWSAny AWS API, through sandboxed scriptsYesNone in the clientIAM policies and condition keys
CloudflareThe whole Cloudflare APIIf grantedOAuth scope choiceCloudflare token scopes
MCP ToolboxTables the database role can seeIf the role canCustom tools, toolsetsDatabase role privileges
FilesystemThe directories you listYes--read-only, --read-only-dirOS file permissions, container mounts
GitRepositories the process can openLocal commits, resets, checkouts--repositoryOS file permissions, container mounts
ssh-mcp-guard (ours)Named actions on your servers over SSHNo: changes need a signed approvalRoot-owned policy on each serversshd forced command on the server

Three rules apply to every server on this list:

  • A local server runs with your permissions. The MCP project's security guidance warns that a server started on your machine can do anything your user can, such as read SSH keys and cloud credentials, unless something sandboxes it.
  • Everything a server returns is untrusted. Issue comments, web pages, log lines and community docs can contain instructions aimed at the model. Claude Code's MCP docs warn that servers fetching outside content carry this risk.
  • Reference servers are examples. The MCP project says its reference servers aren't production-ready. As of 15 Nov 2026 their repository had published 21 security advisories: 8 for Filesystem, 7 for Git, 5 for Fetch and 1 for Everything. Keep them updated, or use a vendor's server where one exists.

How do you sandbox an MCP server?

Limit the credential first, then the server, then the process. The credential matters most because it's enforced by the service, not by code running on your laptop. Give each server its own narrowly scoped credential: a fine-grained GitHub token for named repositories, a Grafana service account with only the permissions its tools list, a Kubernetes view role, a SELECT-only database role, read scopes on Cloudflare and an IAM policy that keeps agent actions read-only. Then turn on the server's own switch (--read-only, --disable-write, read_only = true) and load only the tools you use. Finally, run local servers away from your home directory, in a container or as a separate OS user, with only the directories they need mounted, read-only where possible. Docker's MCP Gateway does this for servers in its catalog: each runs in its own container, and secrets stay out of environment variables. Pin versions instead of running @latest, so new code doesn't arrive unreviewed.

Claude Code adds a few controls of its own:

  • Scopes. Servers can be local (this project, just you), project (shared through .mcp.json in the repository) or user (all your projects).
  • Approval of project servers. Claude Code asks before using servers from a repository's .mcp.json in interactive sessions. In claude -p and Agent SDK runs it can't ask and loads them, so read a repository's .mcp.json before running it headless. disabledMcpjsonServers blocks one everywhere.
  • Output limits. It warns when a tool returns more than 10,000 tokens and caps output at 25,000 by default (MAX_MCP_OUTPUT_TOKENS changes it).

What we tested

We ran the two reference servers most often added by hand on a Mac (macOS 14.5) on 14 Nov 2026 (UTC). Each was installed into a scratch directory and driven with raw MCP JSON-RPC requests (initialize, tools/list, tools/call), against scratch folders and throwaway repositories. Paths are shortened below.

# Filesystem 2.3.0, started with --read-only ws/
tools (10): read_file, read_text_file, read_media_file, read_multiple_files, list_directory,
            list_directory_with_sizes, directory_tree, search_files, get_file_info, list_allowed_directories
write_file ws/b.txt          -> error -32602: Tool write_file not found
read_text_file ws/link/secret.txt  (symlink to a folder outside ws/)
                             -> Access denied - symlink target outside allowed directories

# Filesystem 2.3.0, started with --read-only-dir ref/ ws/
tools (16): the 10 above + write_file, append_file, edit_file, create_directory, move_file, copy_file
write_file ws/b.txt          -> Successfully wrote
write_file ref/notes.md      -> Access denied - read-only directory

# Git 2026.11.11, started with --repository repoA/
git_log repoA                -> OK
git_status repoB             -> Repository path is outside the allowed repository
git_checkout --orphan=x      -> Invalid branch name: cannot start with '-'
git_diff --output=/tmp/x     -> Invalid target: cannot start with '-'

Both servers did what their READMEs say. That's encouraging, but it only covers the cases we tried. The published advisories show how often path checks like these have needed fixing, so run current versions.

Remote or local: which MCP servers are safer?

Neither is safer by default; they fail in different ways. A local server, started over stdio, runs on your computer with your user's rights. If it's malicious, or tricked into misusing a tool, it can reach your files, keys and network. Its tool list doesn't limit that; only a sandbox does. A remote server runs at the vendor, so it can't touch your laptop. It can do only what your OAuth grant or token allows, and you can usually choose those permissions at sign-in, as with Cloudflare, or narrow them later, as with Sentry's skills. Its costs are trust in the vendor's service and your data passing through it. For SaaS you already use (GitHub, Sentry, Cloudflare, AWS), the vendor's remote server with a narrow OAuth grant is usually the better choice. For local resources such as files, Git repositories, clusters and databases, use a local server inside a container or as a separate user, with a scoped credential.

Do you need an MCP server at all?

Often not. A coding agent with a shell can already run gh, kubectl, aws, psql and git with your credentials, and Claude Code's permission rules decide which of those commands it may run. An MCP server earns its place when it adds something the command line doesn't: OAuth instead of a long-lived token on disk, a read-only mode, tool-level switches, or a remote service that keeps work off your machine. Vendors are moving the same way. Microsoft suggests coding agents try the Playwright CLI with skills, and Context7 offers a CLI-plus-skill mode, because tool definitions use up context. If a CLI and a scoped credential already do the job, that's one less server running with your permissions. A quick test: if you'd happily give the agent the same credential on the command line, an MCP server adds convenience, not safety. If you want the agent to have less power than your own login, choose a server whose limits the service itself enforces.

Other MCP servers we looked at

  • Terraform MCP Server (HashiCorp, MPL-2.0, 1.4.1 on 23 Oct 2026): works with the Terraform Registry and HCP Terraform APIs. Worth adding if you write Terraform daily.
  • Docker MCP Gateway (Docker, MIT, 0.44.3 on 26 Oct 2026, marked pre-release like all its releases): runs catalog servers in containers behind one gateway. It's a way to sandbox servers rather than a server itself.
  • Chrome DevTools MCP (Chrome DevTools team, Apache-2.0, 1.13.0 on 11 Nov 2026): controls and inspects a live Chrome browser, with DevTools performance traces. An alternative to Playwright for debugging.
  • Postgres MCP Pro (Crystal DBA, MIT): a restricted mode that allows only read-only transactions, plus performance analysis. Commits continued into August 2026, but its last release, 0.3.0, was on 16 May 2025.
  • Supabase MCP (Apache-2.0, active): the pick if your database is on Supabase.
  • Fetch (reference): fetches web pages. It now refuses private and cloud-metadata addresses by default, but it has had 5 advisories, and every page it fetches is untrusted input.
  • Archived reference servers (GitHub, GitLab, PostgreSQL, Puppeteer, Redis, Sentry, Slack, SQLite and others): moved to the servers-archived repository, whose last commit was on 28 May 2025. Old tutorials still recommend them; use the vendors' servers instead.

For SSH access to servers, which none of the above cover, see our ranking of SSH MCP servers. We read each one's code to see where it enforces its rules.

How we checked this: The reference Filesystem server 2.3.0 (npm) and Git server 2026.11.11 (PyPI), run on a Mac (macOS 14.5) on 14 Nov 2026 (UTC) against scratch directories and throwaway Git repositories, driven with raw MCP JSON-RPC requests. We checked the read-only modes, a symlink pointing outside the allowed directory, and --repository confinement. The other ten servers were reviewed from their repositories, documentation and security advisories on the same day, not run by us. Anything we didn't run ourselves comes from the official documentation listed under Sources.

Give AI agents SSH with a seatbelt

ssh-mcp-guard lets an AI agent read freely but makes every change wait for a signed human approval, enforced on the server. Free and open source.

About ssh-mcp-guard

FAQ

What are the best MCP servers for Claude Code?

For most developers: GitHub's official MCP server, Context7 for up-to-date library docs, and Playwright MCP for browser testing. Then add the vendor server for the systems you run, such as Sentry, Grafana, Kubernetes, AWS or Cloudflare. In Claude Code, add a remote server with claude mcp add --transport http <name> <url> and a local one with claude mcp add <name> -- <command>.

Are MCP servers safe?

Only as safe as the access you give them. A local MCP server runs with your user's permissions, so it can read anything you can, whatever its tools say. Anything a server returns, such as an issue comment, a web page or a log line, can carry instructions aimed at the model. Use official servers, pin their versions, turn on read-only modes and give each one a credential scoped to what it needs.

Where can I find a list of MCP servers?

The official MCP Registry at registry.modelcontextprotocol.io lists published servers. The modelcontextprotocol/servers repository on GitHub now keeps only seven reference servers and points to the registry for everything else. Older reference servers such as GitHub, PostgreSQL, Puppeteer, Slack and SQLite were moved to an archive repository and aren't maintained: use the vendors' own servers instead.

What's the difference between a local and a remote MCP server?

A local server is a program on your computer that the client starts and talks to over stdio, such as the Filesystem or Kubernetes servers. It can reach your files and network. A remote server runs at the vendor, such as GitHub's, Sentry's or Cloudflare's, and you sign in with OAuth or a token. It can only do what that sign-in allows, but your data passes through the vendor's service.

Do MCP servers cost money?

Most servers are free and open source; you pay for the service behind them, if any. On 15 Nov 2026: Context7's free plan allows 1,000 API calls a month and Pro is $10 per seat a month; Sentry has a free one-user Developer plan; AWS charges nothing extra for its MCP server, only for the resources your agents use.

How many MCP servers should I install?

As few as you use every week. Each server's tool definitions take up the model's context, and every extra server is more code running with your permissions. Cloudflare's own figures show the scale: exposing its whole API as one tool per endpoint would take at least 244,000 tokens, which is why its server uses a few code-mode tools instead.

Sources

  1. Model Context Protocol reference servers: README (reference implementations, not production-ready; archived servers)
  2. Security advisories for the MCP reference servers (Filesystem, Git, Fetch, Everything)
  3. MCP documentation: Security best practices (local server compromise, scope minimization)
  4. Claude Code docs: Connect Claude Code to tools via MCP (scopes, .mcp.json approval, output limits)
  5. GitHub MCP Server: README (toolsets, read-only mode, lockdown mode)
  6. Playwright MCP: README (options, isolated mode, security)
  7. Context7: README (tools, disclaimer) and plans
  8. Context7 plans and pricing
  9. Sentry MCP: README (remote and stdio, token scopes, skills)
  10. Sentry pricing
  11. Grafana MCP server: README (RBAC, --disable-write, enforced matchers)
  12. Kubernetes MCP Server: README (TOML configuration, read_only, denied_resources)
  13. Agent Toolkit for AWS: README (AWS MCP Server, IAM condition keys, CloudTrail)
  14. AWS: The AWS MCP Server is now generally available (6 May 2026)
  15. AWS docs: Understanding the MCP Server tools
  16. Cloudflare API MCP Server: README (code mode, OAuth, token costs)
  17. MCP Toolbox for Databases: prebuilt configs and security guidance
  18. Filesystem MCP server: README (allowed directories, --read-only, --read-only-dir)
  19. Git MCP server: README (tools, --repository confinement)
  20. Docker MCP Gateway: README (servers in isolated containers, secrets)
  21. Docker MCP Gateway releases (each marked pre-release)
  22. Git MCP server source: tool annotations (8 read-only, 6 write tools)
  23. Cloudflare's older MCP server repository: the deprecated product servers