Using an MCP client on iPhone: the listed connectors, your own server and the limits

Orvena is an MCP client for the iPhone. It connects the assistant to Notion, Linear, GitHub, Jira & Confluence, Zapier and Home Assistant, or to an MCP server you run yourself, and lets Apple Intelligence, Private Cloud Compute, Orvena's own model or an OpenRouter model call their tools. This guide covers the setup screens by name, the rules a custom server has to meet, how tools fit into a small context window, and the approvals you will see before anything is changed or sent.

What an MCP client on iPhone does

MCP, the Model Context Protocol, is an open standard for connecting services to AI assistants. A service runs an MCP server that lists its tools, each with a name, a description and the arguments it takes, and the client calls a tool when the model asks for it. Orvena is the client, talking to each server over HTTPS directly from your iPhone.

Orvena connects only to remote servers over MCP streamable HTTP. A local stdio server, the kind a desktop app starts as a command, cannot run inside Orvena, so put it behind a remote HTTPS bridge and add that address instead.

The connectors Orvena lists

Open the menu at the top left, then Settings, then Connections (the row titled "Connect your stuff"). Version 2.0.0 lists six services:

The model reads these descriptions to decide which service a request needs. Each server reports its own tools after sign-in, and the row shows the count. Slack, Figma, Canva, Todoist and PayPal are not listed because Orvena cannot meet their sign-in requirements today.

1. Tap Connect on a card. Without Premium Unlock the button reads Premium and opens the purchase screen.
2. For Notion, Linear and Jira & Confluence, the service's own sign-in page opens in a sheet once the server asks for sign-in. Approve access there.
3. For GitHub, a sheet titled "Authorize Orvena" shows a one-time code. Tap "Copy code and open GitHub", paste the code and approve. The sheet reads "Waiting for authorization…" until GitHub confirms.
4. The service moves to the top list, and its status changes from "Connecting…" to its tool count.

Home Assistant MCP on iPhone, and Zapier

Both use a token instead of a sign-in page. In Home Assistant, add the "Model Context Protocol Server" integration, then create a token under your profile, Security, Long-lived access tokens. In Orvena, tap Connect on Home Assistant, enter the address under "Where does it live?", such as https://your-home:8123/api/mcp, and paste the token.

Orvena requires HTTPS here even on your own network, so it will not send the token to an http:// address, such as http://homeassistant.local:8123 or a local IP address over http://. With an http:// address the sheet's Connect button does nothing. Home Assistant needs a valid TLS certificate or a trusted HTTPS reverse proxy.

Zapier asks only for the token: at mcp.zapier.com, create a server, choose its apps, and copy the connection token from the Connect tab.

How to add your own MCP server

The last card on Connections, "Your own server", opens "Custom server". The URL must use https:// with a host name and contain no user name, password or #fragment. The server must speak MCP streamable HTTP, answering with plain JSON or an SSE stream, and support protocol version 2025-11-25, 2025-06-18 or 2025-03-26. Redirects must stay on the same origin, and a server that reports no tools is not saved.

1. Fill in Name and the address field, whose placeholder is https://server.example/mcp.
2. If the server uses an API key or token, paste it into "Bearer token (optional)". Left empty, Orvena tries open access, then sign-in.
3. If an OAuth server does not say which scopes it needs, enter them in "OAuth scopes, space-separated (optional)".
4. In "What can this server help with? (optional)", write a sentence about what it is for.
5. Tap "Add server". It reads "Checking…" while Orvena connects and lists the tools.

Orvena starts OAuth only after an HTTP 401, 403 or WWW-Authenticate challenge, and the authorization server must support PKCE with S256.

Orvena uses the purpose line only to decide when to load the server. Without one, it builds each server's summary from up to four tool descriptions, cut to 320 characters. If any tool is missing a description, the row shows "Add a purpose so Orvena can choose this server reliably.", and if none has one, the server is used only when you name it. Fix that with "Describe purpose" in the row's ••• menu.

What needs Premium Unlock

All connections need it, the six listed services and any server you add yourself. Premium Unlock is a one-time $14.99 purchase with no subscription, and it also adds Private Cloud Compute on iOS 27, cloud models on your own OpenRouter key, voice conversations and scheduled tasks. Apple Intelligence and Orvena's downloaded model are free, but without the unlock they answer without connector tools.

Using MCP with Apple Intelligence and the other engines

All four engines on the Models screen get the connector tools. For Apple Intelligence and Private Cloud Compute, Orvena converts each MCP tool's JSON schema into the schema format of Apple's Foundation Models framework. If a tool's schema uses something the framework cannot express, Apple's models get that tool with no arguments, so use Orvena's model or OpenRouter for it. An OpenRouter model listed without tool support is the exception: "This model can't use tools, so Orvena will answer without them."

Automatic, the default engine setting, never picks Private Cloud Compute or OpenRouter on its own.

How tool lists and results fit a small window

An 8,192-token window has little room for dozens of tool definitions, so connector tools never go to the model all at once. The model first calls load_connectors to pick one or two services from a list of the connected services, each described in at most 180 characters (8,192 characters in all). Then load_tools picks one to eight operations from them. Only those become callable.

Before every model round, Orvena adds up the instructions, the loaded tool schemas and 900 tokens of room. If that exceeds the window minus 512 tokens, it drops connector tools, newest first, except ones the request still has to call, and then the oldest skill. If the needed operation still does not fit, the turn stops with "The required connector operation is too large for this model's context window."

Tool descriptions are cut to 500 characters, and results are marked as untrusted data, which tells the model to treat text from a page or issue as content, not instructions. A result over the budget is compacted: pagination fields come first, links and long hashes go, objects in a list keep at most 12 fields, and long strings are shortened, with a note telling the model so.

In long conversations Orvena keeps whole recent turns within the engine's window and drops the oldest ones. It does not summarize. When even that does not fit, Apple Intelligence shows "This conversation is too long for Apple Intelligence. Start a new chat, or switch to Orvena's downloaded model in Models."

Approvals before anything changes or leaves

Orvena classifies each MCP tool from the hints its server declares: read-only tools show "Reads data without changing it.", destructive ones "Deletes, cancels, or performs an irreversible action.", and the rest "Creates or changes data."

• Read and write tools ask the first time. The sheet, headed "APPROVE THIS ACTION", has "Don't allow" and "Allow", with "Allow once" or "Always allow" in its menu.
• Destructive tools ask every time, with "Allow once" only.
• The sheet lists the exact arguments, with tokens, passwords and other credential-like fields shown as [REDACTED CREDENTIAL], and says where they go, for example "Sends the data needed for this action to Notion."
• If the same turn already read something personal, such as your calendar or another connector's result, the approval is titled "Send data to" and the service's name. After Health data it asks every time.

Private Cloud Compute asks once, when you choose it on Models, and says that each answer sends "your conversation, and the tool results it needs, to Apple's servers." Connector approvals then show "Notion → Private Cloud Compute" as the destination. OpenRouter asks before the first request in a conversation, and again with "Use OpenRouter with sensitive data?" once a connector result is in the history. Saved grants are under Settings, Data sharing, where you can revoke them.

Where the data goes

Connector traffic goes from the iPhone straight to the service, with no Orvena server in between. The service and its sign-in server receive the arguments you approved, the content they return and the sign-in traffic. Tokens stay in the iPhone's Keychain, on this device only. With Apple Intelligence or Orvena's model the results are read on the phone; with Private Cloud Compute the results an answer needs go to Apple, and with OpenRouter to OpenRouter after you consent. Settings, Activity keeps up to 30 days of tool names, outcomes and destinations, never arguments. Disconnect removes the local credentials and tools, and the confirmation reminds you to revoke access on the service as well, for example "You may still need to revoke Orvena in Notion's account settings." The broader picture is in what makes an AI assistant private.

Troubleshooting

Common questions

Can I run an MCP server on my iPhone?

Not inside Orvena. Orvena is an MCP client that connects to remote servers over HTTPS using MCP streamable HTTP. A local stdio server cannot run inside Orvena, so run it elsewhere behind an HTTPS bridge and add that address under Your own server.

Can Apple Intelligence use MCP tools?

In Orvena it can. On iOS 27, Orvena gives Apple's on-device model and Private Cloud Compute the tools of the servers you connect, converted to the format of Apple's Foundation Models framework. At any moment it has one or two services and at most eight of their operations loaded, which keeps the tool definitions small enough for the on-device window.

How do I connect Home Assistant to an AI assistant on iPhone through MCP?

Add the Model Context Protocol Server integration in Home Assistant and create a long-lived access token under your profile, Security. In Orvena, tap Connect on Home Assistant, enter the HTTPS address ending in /api/mcp, and paste the token. Plain http addresses are refused, including on your home network.

Can I use the Notion MCP server on iPhone?

Yes. Notion is one of Orvena's six listed connectors. Tap Connect under Settings, Connections, sign in to Notion and approve access. Orvena uses Notion's hosted server at mcp.notion.com and asks before a Notion tool first runs. Connections are part of Premium Unlock.

Does every MCP action ask for approval?

Destructive actions ask every time. Tools that read, create or change data ask the first time, and you can choose Always allow. That grant is listed under Settings, Data sharing, Allowed tools, where you can revoke it.