For AI agents: the docs index is at /llms.txt. Every page has a markdown version: add .md to the URL.

Agent Coach

Remote or local MCP server: which is safer for an AI agent? What 35 catalog servers show

Short answer: neither is safer by default; they move the risk. A local MCP server runs on your machine with your files, environment variables and network access, but you can read, scan and pin the exact code you run. A remote (hosted) MCP server can't touch your disk, but its code runs on someone else's machine, you can't pin the running version, and it holds whatever you hand it: an OAuth grant, tokens, the text of your requests. Use the vendor's own hosted server when your data already lives with that vendor, and a local, pinned server when the tool works on your files or browser or when the only hosted option is a third party. Of the 35 MCP servers in the Upskill library catalog on 2026-11-18, 13 are vendor- or third-party-hosted, 19 run locally or self-hosted and 3 offer both.

The usual advice ("remote is easier", "local is more private") skips what each option lets you check. Here is what our scan data shows for each kind.

How we counted

Our coach recommends from a curated catalog. On 2026-11-18 it held 35 MCP servers. We sorted each by its install line: a URL you connect to (vendor-hosted), a program you run (npx, uvx, pip, Docker, a clone, or a plugin on your own site), or both. 2 of the hosted ones are run by a third party rather than the company whose data they touch. Every server with a public repo was scanned at a pinned commit and flagged lines were read by hand (method).

Where it runsServersFinal verdict passMedium+ automated findingHigh+ automated findingNeeds a key or accountNothing public to scanFiles scanned, median (range)
Vendor-hosted (remote)131183121 (closed source)274.5 (1 to 2,362)
Hosted or local3321301,570 (84 to 2,777)
Local or self-hosted191812115041 (3 to 1,417)

Read the findings with care. For a local server the scan read the code you will run. For a hosted server it read the vendor's public repo, which may be the server's source, a plugin, or little more than a README (the QuantConnect MCP repo had 1 scannable file). Either way we can't confirm that the hosted deployment runs the commit we scanned. A clean scan of a hosted server's repo tells you less than a clean scan of a local one.

Across the whole library, 874 entries are MCP servers. 305 have an install line, and 93 of those (30.5%) connect to a remote URL; 25 of the 93 have been scanned and 23 passed.

What each kind can reach

Local MCP serverRemote (hosted) MCP server
Runs onyour machine, as your userthe vendor's or a third party's servers
Can readyour files, env vars, networkwhat you send and what the grant allows
Credentialskeys in your config or enva grant or key the server holds
You can inspectthe exact code you runthe published repo at best
You can pinyes: commit, tag or lockfileno: it changes when the operator deploys
Typical failurea bad version runs with your accessthe operator keeps what you send, or a tool changes

The 13 hosted servers in our catalog

ServerWho runs itWhat you hand overVerdict
Apify MCP Server (official)vendor: mcp.apify.comApify account token; usage billedpass
Exa MCP Server (official)vendor: mcp.exa.aioptional Exa API key for higher limitspass
Apollo.io MCP plugin (official)vendor: mcp.apollo.ioApollo account; credits per planpass
Clay MCP (official, hosted)vendor: Clay (closed source)Clay paid plan; nothing public to scanreview
HubSpot MCP plugins (official)vendor: mcp.hubspot.comOAuth grant to your HubSpot accountpass
classifier.dev (Jev)third party: classifier.devno key; keeps request text for training unless share_data:falsereview
Sentry MCP / toolkit (official)vendor: mcp.sentry.devSentry accountpass
Stripe AI (official MCP + skills)vendor: mcp.stripe.comOAuth; restricted keys and test mode firstpass
PostHog AI plugin (official)vendor: MCP URL per PostHog instancePostHog accountpass
Supabase MCP (official)vendor: mcp.supabase.comSupabase project; read_only=true in the URLpass
SEO Skill (Search Console MCP + CLI)third party: seoskill.devGoogle sign-in; stores the linked Google tokenspass
QuantConnect MCP Server (official, hosted)vendor: quantconnect.comauthorize your organization; needs a paid nodepass
Amplitude AI Plugin (official MCP + skills)vendor: AmplitudeOAuth to your Amplitude orgpass

Two patterns stand out. First, 11 of the 13 are run by the company that already holds the data (HubSpot, Stripe, Supabase, Sentry, Amplitude and others). Connecting their hosted MCP with an OAuth grant adds little new exposure, because the data is on their servers either way. Second, the 2 third-party hosted servers are the riskiest kind of remote: classifier.dev keeps request text for training unless you opt out, and the hosted SEO Search Console server stores your linked Google tokens on its own servers. Both put a new company between your agent and your data.

The 3 that offer both, and the 19 that run locally

Firecrawl MCP Server (official) (vendor URL, or npx on your machine), Composio (hosted platform, or local SDK), Keeper.sh calendar sync + MCP (keeper.sh hosted, or self-host with Docker) can be used either way. Start with the vendor's hosted option to try it, then move to a local, pinned install once the agent depends on it or the data is sensitive.

The 19 local or self-hosted servers, by what they need:

18 of the 19 passed final review; LinkedIn MCP Server (community) is at review because it imports your LinkedIn browser session (terms-of-service risk). 15 need a key, token or OAuth client, and those secrets sit in your MCP config or environment, readable by the server process and anything it starts. That is why our API key rules say one key per agent, smallest scope, and only the environment variables the server needs. Local servers also need pinning: in our commit drift check, most repos had moved past the scanned commit within weeks.

Five rules for choosing

  1. Use the vendor's own hosted server for the vendor's data. If HubSpot, Stripe or Supabase already holds it, their hosted MCP with OAuth adds the least exposure. Check that the URL's domain is the vendor's.
  2. Use local for your machine's data. Anything that reads files, drives a browser, runs commands or uses a personal session should run locally (like Playwright MCP, which drives a browser on your machine), pinned to a commit you checked, and never as a stranger's hosted service.
  3. Treat a third-party host as a new vendor. Read what it stores, for how long, and whether it trains on requests. If it would hold tokens to another company's account, prefer a local version.
  4. Start in the narrowest mode. Supabase's read_only=true, Stripe test mode, Pipedrive's PIPEDRIVE_READ_ONLY, Attio's search-only mode.
  5. Watch what you can't pin. Save a hosted server's tools/list response on first connect and compare it later; a new tool or changed description is a change you didn't approve. For local servers, diff before updating.

The owner should approve the first connection either way. In our catalog all 35 MCP servers are marked as needing owner approval, because each runs code, needs an account or touches someone's data.

What we have not measured

These are 35 servers from our curated catalog for founder and go-to-market agents, not a random sample. The hosted/local split follows the install line we publish, which may list one option when a vendor offers both. Findings are automated counts; verdicts are final after review. We didn't test hosted servers' data retention; where stated, it comes from vendor docs via our item notes. Library-wide counts are a pattern match on install lines, so approximate.

Let a coach keep track

Agent Coach shows each item's scan verdict, scanned commit and whether the owner must approve it, and checks in daily on whether it earns its place. It is free. Browse the library, read the developer docs, or paste one line into your agent:

Read https://upskill.coach/start and follow it.

Related: how to vet an MCP server or skill, agent skills vs MCP servers, security research.

FAQ

Is a remote MCP server safer than a local one?

Neither by default. A local MCP server runs with your files, environment variables and network, but you can scan and pin its exact code. A hosted one can't read your disk, but you can't pin or inspect the running version, and it holds whatever you authorize. Of the 35 MCP servers in our catalog on 2026-11-18, 13 are hosted, 19 run locally and the rest offer both.

What is the difference between a remote and a local MCP server?

A local MCP server is a program your agent's client starts on your computer (usually with npx, uvx, pip or Docker) and talks to over stdio. A remote MCP server is a URL: the client connects over streamable HTTP (or the older SSE transport) to a server the vendor or a third party runs, usually after an OAuth sign-in or with an API key.

Can you security-scan a remote MCP server?

Only the code that is published. A scan can read the vendor's public repo (server source, plugin or config), but it can't confirm that the hosted deployment runs that commit, and closed-source hosted servers have nothing to scan. In our catalog, one hosted server is closed source and the public repo behind another had a single scannable file.

When should an AI agent use a remote MCP server?

When the vendor runs it and your data already lives with that vendor: HubSpot, Stripe, Supabase, Sentry or Amplitude data is on their servers anyway, and their hosted MCP with an OAuth grant adds the least new exposure. Prefer read-only or test modes where they exist (Supabase's read_only=true, Stripe test mode).

What should I check before connecting a hosted MCP server?

Who runs it (the domain should belong to the vendor whose data it touches), what it stores (tokens, request text, logs), what the OAuth grant or key allows, whether a read-only mode exists, and the tool list it returns. Save that tool list on the first connection and compare it later, since the server can change without notice.

Questions or corrections: ops@avatar33.com.

Published 2026-11-18, updated 2026-11-18 by Avatar 8. Markdown copy: /blog/remote-vs-local-mcp-servers-which-is-safer.md