Which AI apps support custom MCP servers (2026 comparison)

A custom MCP server is any Model Context Protocol server you point an AI app at yourself, by URL, rather than one from a built-in directory or app store the vendor curates. Remote means the server is reachable over HTTP — a Streamable HTTP or legacy SSE endpoint sitting somewhere reachable by URL, as opposed to a local (stdio) server the app launches as a subprocess on your machine. This page covers remote only: every row below is "does this app let me add my own server by URL," not "does this app support MCP at all" (several also support local servers, a separate config shape).

Every claim below comes from that app's own documentation, read on the date shown. Nothing here is a claim about how well an app works with any particular server — including ours. Where a guide's own text uses a generic example, this page describes that example the same way the guide does: a standard remote MCP server, connected the way that app connects any standard remote MCP server.

How to read this table

• Where you add it — the UI path or the config file key, per that app's docs.
• Remote transport — Streamable HTTP, SSE, or "bridge" when the app is really configuring a harness underneath it (GLM) rather than the app itself.
• OAuth sign-in — whether the app can complete an OAuth flow with the server on its own. "Manual" means you supply a client ID/secret yourself; "no IDE flow" means the app has no OAuth UI at all and needs a proxy.
• Plan / availability notes — any gate the vendor's own docs state: paid plan, admin/maker role, region, or platform restriction. "Not documented" means the source guide didn't find a stated gate, not that the app is confirmed gate-free.
• Guide link — the full walkthrough, config examples, and troubleshooting FAQ for that app, each independently dated.

Guides span three verification windows: seven guides from August 15–24, 2026; fourteen from September 15, 2026; and eleven verified or re-verified September 28, 2026. All facts below are as of September 2026 — check the linked guide's own lastVerified date before relying on any single cell.

Chat apps

App	Where you add it	Remote transport	OAuth sign-in	Plan / availability notes	Guide
ChatGPT	Settings → Apps → Advanced Settings → Create App (also Settings → Security and login → Developer mode)	Custom app with OAuth, Developer Mode required	Yes	Paid plan required (Plus, Pro, Business, Enterprise, Edu) — not Free. Web only. Business/Enterprise/Edu need an admin to enable custom MCP first. Write tools are limited on Plus/Pro even though lookups work.	/guides/add-mcp-server-to-chatgpt
Claude	Customize → Connectors → Add custom connector	Remote MCP over HTTPS	Yes (or static request headers)	Available on every plan including Free — Free is capped at one custom connector at a time. Runs from Anthropic's cloud, so add once on web/desktop and it syncs to mobile on next sign-in.	/guides/add-mcp-server-to-claude
Gemini (consumer app)	gemini.google.com → Settings → Connected Apps → Custom apps → Add a custom app (web only)	Custom app added by URL	Yes — a browser consent screen opens when you add the app	Must be 18+, in the US, on a personal (not work/school) Google account, with Keep Activity on; English only. Added on web, then available in the mobile app.	/guides/add-mcp-server-to-gemini
Grok	grok.com/connectors → New Connector → Custom	Custom connector by URL, automatic tool discovery	Completed in-app if the server requires it (per xAI docs)	Every plan including Free. Business/Enterprise need a team admin to provision the connector first. Server must be reachable over the public internet — Grok rejects localhost and private IP ranges.	/guides/add-mcp-server-to-grok
Perplexity	Account settings → Connectors → + Custom connector → Remote	Streamable HTTP or SSE, chosen per connector	Yes — one of three auth options (None, API Key, OAuth 2.0)	Pro, Max, and Enterprise (shipped March 2026). A separate, older Perplexity help page still says remote MCP is "coming soon" — that page is stale; follow the custom-connector article instead.	/guides/add-mcp-server-to-perplexity
Mistral Vibe (formerly Le Chat)	Connectors page (under Work) → + Add Connector	Custom MCP Connector by URL, transport not specified by name — auth auto-detected	Yes, auto-detected — None, OAuth 2.1, or Bearer/Basic all work without picking a mode	No stated plan gate in Mistral's docs — requires administrator access, and on Free/Pro/Student the account owner is the admin by default. Dynamic tool discovery is a documented current limitation: the tool list doesn't auto-update when the server changes.	/guides/add-mcp-server-to-mistral-vibe

IDEs / editors

App	Where you add it	Remote transport	OAuth sign-in	Plan / availability notes	Guide
Cursor	Customize (sidebar) → MCPs, or .cursor/mcp.json / ~/.cursor/mcp.json	Native Streamable HTTP, no proxy	Yes, automatic on first connection	Not documented as plan-gated. Customize panel requires v3.9+ (June 22, 2026); older builds edit mcp.json directly.	/guides/add-mcp-server-to-cursor
VS Code	Command Palette → MCP: Add Server, or .vscode/mcp.json (root key servers, not mcpServers)	Native HTTP	Yes, via an Auth CodeLens	Not documented as plan-gated directly, but a Copilot org policy can hide MCP entirely. Tools only surface when Chat is in agent mode.	/guides/add-mcp-server-to-vs-code
JetBrains IDEs	Settings → Tools → AI Assistant → Model Context Protocol (MCP) → Add	Remote HTTP, requires a 2025.3+ build; bearer headers only	No IDE OAuth flow — AI Assistant has none; use the mcp-remote proxy for OAuth-protected servers	AI Assistant has Free-tier exclusions per JetBrains' licensing docs. Junie (the JetBrains agent) does have its own OAuth Authorize flow and can reuse the same server configs.	/guides/add-mcp-server-to-jetbrains
Zed	Agent Panel settings menu, or settings.json → context_servers → url	Native, tokens stored in the system keychain	Yes, automatic — but only when no Authorization header is configured (one skips the other)	Not documented as plan-gated. Open Zed issue as of August 2026: tools from OAuth-authenticated servers may not appear in an external agent driven through ACP.	/guides/add-mcp-server-to-zed
Windsurf (Devin Desktop)	Cascade: ~/.codeium/windsurf/mcp_config.json → mcpServers → serverUrl; Devin Local: .devin/mcp_config.local.json → url	HTTP server entry	Yes — Devin Local via devin mcp login or on first use	Enterprise gate: MCP ships off by default on Enterprise. Two separate agents (legacy Cascade vs. the now-default Devin Local) read different config files and won't see each other's servers.	/guides/add-mcp-server-to-windsurf
Cline	MCP Servers icon → Remote Servers tab, or cline_mcp_settings.json	Streamable HTTP — requires "type": "streamableHttp" explicitly; omitted defaults to legacy SSE	Yes, for OAuth-capable servers (no headers entry needed) since v4.1.7 added pre-registered OAuth client support; earlier versions only documented static header auth	Not documented as plan-gated.	/guides/add-mcp-server-to-cline

CLIs / agents

App	Where you add it	Remote transport	OAuth sign-in	Plan / availability notes	Guide
Claude Code	claude mcp add (add --scope user for every project), or .mcp.json → mcpServers with "type": "http"	Streamable HTTP	Yes, browser OAuth via /mcp	Not documented as plan-gated. Config entries missing type are skipped (with an error in v2.1.202+) rather than silently failing.	/guides/add-mcp-server-to-claude-code
OpenAI Codex	codex mcp add, or ~/.codex/config.toml → mcp_servers.<name> → url + optional bearer_token_env_var	Streamable HTTP	Yes, via codex mcp login <name>; bearer token is the alternative	Same config shared by the CLI, IDE extension, and ChatGPT desktop app — add once, available in all three. A project-local config.toml only loads for trusted projects.	/guides/add-mcp-server-to-codex
Gemini CLI	gemini mcp add --transport http ... -s user for global, or settings.json → mcpServers → httpUrl (not url — that key means SSE)	Streamable HTTP via httpUrl	Yes, via /mcp auth	Stopped serving consumer Google accounts (Free, AI Pro, Ultra) on June 18, 2026. Works with Gemini Code Assist Standard/Enterprise licenses and paid Gemini API keys.	/guides/add-mcp-server-to-gemini-cli
Qwen Code	~/.qwen/settings.json or .qwen/settings.json → mcpServers → httpUrl (same url-vs-httpUrl trap as Gemini CLI)	Streamable HTTP via httpUrl	Yes, automatic dynamic client registration via /mcp	Not documented as plan-gated.	/guides/add-mcp-server-to-qwen-code
OpenCode	opencode.json or ~/.config/opencode/opencode.json → mcp → type: "remote" + url	Remote HTTP	Yes, automatic discovery	Not documented as plan-gated. type is required by the config schema — an entry without it silently fails to load.	/guides/add-mcp-server-to-opencode
Kiro	.kiro/settings/mcp.json (project) or ~/.kiro/settings/mcp.json (global); a custom agent's own mcpServers outranks both	URL-only entry	Yes, automatic — attempts dynamic client registration, falls back to a default client	Nothing in Kiro's pricing tiers gates MCP (free tier works the same as paid). Real gates: no MCP in Kiro Mobile, and an enterprise admin can disable MCP or restrict it to an allow-list.	/guides/add-mcp-server-to-kiro
GitHub Copilot CLI	/mcp add, or ~/.copilot/mcp-config.json → mcpServers → type: "http" + url + headers	Streamable HTTP / SSE	Yes, in current releases — Client ID Metadata Document (CIMD) support landed in CLI v1.0.83 (Sept 4, 2026); GitHub's setup page itself only documents static headers	Not documented as plan-gated. Project-scoped servers (.mcp.json / .github/mcp.json) don't follow you across repos — use the user-level file for that.	/guides/add-mcp-server-to-github-copilot-cli
Warp	Settings → Agents → MCP servers, or ~/.warp/.mcp.json (global) / .warp/.mcp.json (project) → mcpServers → url	Native, tokens stored on-device	Yes, for local agents — browser flow, once per machine	Not documented as plan-gated. Cloud agents are a hard limitation: a url server that requires OAuth with no Authorization header cannot be used by cloud agents at all — use a local agent, or a managed installation referenced by warp_id.	/guides/add-mcp-server-to-warp
Goose	Desktop: Extensions → Add custom extension; CLI: goose configure → Add Extension → Remote Extension; or ~/.config/goose/config.yaml → extensions → type: streamable_http (underscore, not hyphen) + uri	Remote HTTP (streamable_http)	Yes, when the server supports Dynamic Client Registration or a Client ID Metadata Document; otherwise manual client_id (and client_secret_key) on the extension entry	Not documented as plan-gated.	/guides/add-mcp-server-to-goose
LM Studio	Chat sidebar → Program tab → Install and Edit mcp.json → mcpServers → url	Remote HTTP	Yes, since v0.4.10 (MCP support itself since v0.3.17)	Not documented as plan-gated (LM Studio is local/free). Caveat isn't the app but the model: only models with native tool-calling support (marked with a hammer badge) reliably call MCP tools; smaller models can get worse after connecting one.	/guides/add-mcp-server-to-lm-studio
Raycast	Run the Install MCP Server command → fill in name, Transport: HTTP, URL, OAuth Type: Dynamic	Remote HTTP	Yes, OAuth Type Dynamic	MCP is Pro-exclusive as of August 2026 (labeled "Tier: Pro Exclusive" in Raycast's manual across Mac, Windows, iOS). Free tier gets 50 trial AI messages, then is pointed at Pro. Add separately on each platform — installing on macOS doesn't carry to iOS automatically.	/guides/add-mcp-server-to-raycast

Automation / other

App	Where you add it	Remote transport	OAuth sign-in	Plan / availability notes	Guide
Microsoft Copilot Studio	Agent → Tools → Add a tool → New tool → Model Context Protocol (onboarding wizard)	MCP onboarding wizard (Streamable HTTP)	Yes, in three modes: Dynamic discovery (server supports OAuth 2.0 DCR with a discovery endpoint), Dynamic, or Manual (you supply client ID/secret)	Maker/admin flow, not an end-user chat setting — needs a Copilot Studio user license or the Copilot Studio authors role, plus a tenant with Copilot Credit capacity provisioned. Generative orchestration must be turned on for the agent or the server's tools won't be used. Trial licenses can build/test but not publish.	/guides/add-mcp-server-to-copilot-studio
n8n	Add an MCP Client Tool node as a tool on an AI Agent node; set its Endpoint and Server Transport	Remote HTTP: SSE or Streamable HTTP, per the server (node v1.2+ defaults to HTTP Streamable; v1 only had an "SSE Endpoint" field)	Yes, two ways — pick a server from n8n's built-in MCP registry (sign-in creates the credential for you), or add the node manually and choose OAuth2 as the Authentication option (alongside Bearer, Header, and None)	Not documented as plan-gated — this is a workflow-builder node, available wherever n8n's AI Agent node is. This is the opposite direction from the separate MCP Server Trigger node, which makes n8n itself an MCP server.	/guides/add-mcp-server-to-n8n
Open WebUI	Admin Settings → Integrations → External Tool Servers (admin-only)	Remote Streamable HTTP only	Yes, per-user OAuth 2.1, or a bearer token	Registration is admin-only, unlike OpenAPI tool servers which users can self-add if granted permission — once registered, individual users authenticate to it themselves. Self-hosted: a Dockerized Open WebUI reaching a server on the host machine needs host.docker.internal, not localhost.	/guides/add-mcp-server-to-open-webui
GLM (Z.ai)	Not the model itself — configured on whatever harness runs it (Claude Code, OpenCode, Cline)	Harness-level — MCP is not configured on the model at all	Whatever the harness supports (see that harness's row)	Swapping to GLM changes nothing about the MCP connection; it's entirely the harness's config. Included here because "does GLM support MCP" is a common question with a "wrong question" answer.	/guides/add-mcp-server-to-glm

Caveats

• Docs lag releases. More than one guide behind this table exists because a vendor's own help page was stale or contradicted its own changelog — Perplexity's "coming soon" page next to a shipped feature, and Cline's OAuth behavior living only in GitHub release notes, not the docs page. Read the linked guide's sources, not just this table, before making a decision that depends on a specific behavior.
• Recheck before relying on any cell. MCP support across these apps changed multiple times within 2026 (Gemini CLI's consumer-account retirement, Raycast's move to Pro-exclusive, Windsurf's rename and default-agent switch, Mistral's rename to Vibe). Each guide's lastVerified date is the last confirmed-accurate date, not a guarantee it's unchanged today.
• "Not documented" is not "unsupported." Several plan/OAuth cells above say the source guide found no stated gate. That means the vendor's own docs didn't call one out at the time of verification — it is not the same as a vendor statement that no gate exists.
• This table doesn't rank or recommend. It restates what each vendor documents about their own product. Choosing among these apps depends on factors this table doesn't cover — pricing beyond the MCP-specific gates listed, existing workflow, and non-MCP features.