Which AI assistants support custom MCP connectors?
The Model Context Protocol lets you point an AI assistant at your own data. Which assistants actually let you add a server of your choosing is harder to find out than it should be, so here is the matrix, checked in July 2026, including the ones where the answer is no.
The matrix
| Assistant | Custom MCP? | How |
|---|---|---|
| Claude (Desktop, web, Code) | Yes | Settings → Connectors → Add custom connector. Free tier gets one. |
| ChatGPT | Yes | Developer Mode / custom connectors, paid tiers, web. |
| Perplexity | Yes | Custom remote connectors on Pro and Max. |
| Grok | Yes | BYO MCP connector on web, iOS and Android. One of the few mobile surfaces. |
| Mistral Le Chat | Yes | Custom MCP connectors, also exposed in Mistral Studio. |
| Local clients (LM Studio, AnythingLLM, Cline, Continue, Cursor, LibreChat, Msty, Cherry Studio, 5ire) | Yes | stdio, usually via a bridge. Long-lived tokens rather than a browser OAuth flow. |
| Gemini CLI | Yes | Config file. Developer surface, not the Gemini app. |
| Gemini Enterprise | Yes, admin only | Custom MCP data store. Requires Streamable HTTP; rejects the legacy SSE transport. |
| Microsoft 365 Copilot | Maker / admin only | Via a Copilot Studio agent or the Agent Store. No end-user "paste a URL" path. |
| Gemini app (consumer) | No | No custom-connector interface exists. |
| Copilot (consumer) | No | No extensibility surface. |
| Meta AI | No | Meta publishes MCP servers but ships no MCP client. |
The transport catch
There are two remote transports. HTTP+SSE came first and was deprecated in March 2025; Streamable HTTP replaced it, and a stateless revision of that spec finalises on July 28, 2026. The distinction is not academic: Gemini Enterprise rejects SSE outright, and the newer connector surfaces expect Streamable HTTP. If you are choosing an MCP server to connect, check which transport it actually serves. If you are building one, serve Streamable HTTP and keep the SSE endpoint as an alias so existing connectors keep working.
What to ask of any MCP server you connect
Connecting an MCP server hands an assistant real access to real data. Three questions worth asking before you approve the consent screen:
- Can you see what the assistant did afterwards? A tool-call log, with the inputs it sent, is the difference between an audit trail and a promise.
- Can you revoke it, and does revocation actually take effect on the next call?
- Does anything it writes need your approval first, or does it write straight through?
Calmara is an MCP server for your own tasks, calendar, notes and AI memory, and it answers those three with an activity log, revocable tokens with per-scope consent, and a pending-review queue. Connector URL: https://api.calmara.app/mcp. See the trust page for where the data lives, or the MCP security risks nobody mentions for why the questions matter.
Checked July 2026. Assistant capabilities change quickly; if you find a row that is out of date, email support@calmara.app and it will be corrected.