Apple Messages

@mgcrea/mcp-apple-messages

Model Context Protocol server for Apple Messages — iMessage, SMS and RCS — on macOS.

Unofficial. Not affiliated with Apple. It reads the message store already on your Mac.

The surface with no second lane

Every other server in this family has two ways in and picks the faster one. This one has exactly one, and that is measured rather than assumed: every read through Messages' scripting dictionary fails. It answers Application isn't running even while NSRunningApplication reports it running, because it lives as a windowless background process that declines to wake for a script.

So Full Disk Access is mandatory, not an upgrade. Without it this server does nothing at all — there is no degraded mode to fall back to.

Apple Events is still here, for exactly one verb. send is the only command in the dictionary that works, which makes this the one surface where Apple Events is a write lane and nothing else — and therefore the one where the write gate is also a permission boundary. With APPLE_MESSAGES_ALLOW_WRITES off, no Apple Event is ever sent and no Automation grant is ever requested. There is no read to leak through the gate, because there is no read.

One message in thirty-two is invisible to SQL

chat.db keeps message text in two places, and the one you would reach for is the _incomplete_ one. Measured on a real 97,416-message store:

The blob is an NSArchiver _typedstream_, and no amount of SQL reaches inside one. A server that selects text returns nothing for one message in thirty-two, silently and with no error to notice.

And that ratio is a historical average. Measured through this server against a live store: 2016 through 2025 are ~99% plain text, and then it stops — from March 2026 every message is blob-only. Apple appears to have dropped the column between late February and late March. So the decoder is not handling an edge case; it is the only way to read anything recent, and a server without one would report that the conversation stopped in February.

This package decodes it. The decoder was written against ground truth — NSArchiver still ships on macOS 26 and is what wrote those blobs, so archiving a known string produces a fixture nobody has to guess at — and then validated against the whole store using text as an oracle on 94,043 labelled rows:

100.000% agreement. 97,094 of 97,094 blobs decoded, none failed. ~2 ms per thousand.

decodeAttributedBody() is exported from the package root.

Names come from Contacts

chat.db records a correspondent as +15551234567 and nothing else, so this server depends on @mgcrea/mcp-apple-contacts to turn that into a name. Every correspondent carries both, plus a resolution status.

`unknown` is normal, not an error. Measured against a real address book: 97.6% of the last year's messages resolve, but only ~84% of the twenty-five busiest correspondents — so roughly one in six of the people you talk to most has no contact card. Code that treats an unresolved handle as a failure will be wrong several times on any real inbox.

If the Contacts permission has not been granted, resolution is skipped and handles come back raw. That is a capability downgrade reported through diagnostics, never a throw.

The lookup runs with the Contacts package's defaults. The APPLE_CONTACTS_* settings configure the Contacts server, not this one: setting them here changes nothing, and APPLE_MESSAGES_RESOLVE_CONTACTS is the only switch Messages reads for it.

Tools

Read: diagnostics, list_chats, list_messages, search_messages, get_message, count_messages.

count_messages is the arithmetic lane: totals with a sent/received split, and groupBy for counts per day, month, chat, handle or direction. It aggregates in SQL over every match, so "who do I message most" costs one small result rather than a page of conversations tallied by hand — and grouping by handle spans the several chats one person can occupy, which the per-chat totals on list_chats cannot add together. Metadata only, deliberately: message bodies from 2026 onward live in an archived blob that SQL cannot read, so a text filter here would return a number that looks right and silently omits the recent half of the store.

Opt-in read: find_codes, behind APPLE_MESSAGES_ALLOW_CODES — see Configuration. Off by default, and not covered by the write gate.

Write: send_message, and that is the whole dictionary. sdef lists three commands — send, login and logout — and the other two would sign the user out of iMessage on every device they own. There is no edit, delete, mark-as-read or reaction verb to expose, so everything this server can show you, it cannot change.

send's direct parameter is typed file OR text, and both forms ship in two lanes with different bounds. attachmentId forwards a file that is already in this Mac's Messages store — the same guid save_attachment takes, through the same "inside ~/Library/Messages, or refuse" boundary — so it reaches no arbitrary path and needs no flag beyond ALLOW_WRITES. filePath names any local file, which is an exfiltration primitive whose blast radius, unlike the text form's, is not bounded by what the model can say; it exists as a parameter only when APPLE_MESSAGES_ALLOW_FILE_SEND is on, and it is off by default. There is no directory confinement on it on purpose — see docs/messages.md for why one would be a speed bump wearing a boundary's clothes.

One call sends one thing, so a captioned photo is two calls.

Sending, and how it reports what it sent

Prefer a chatRef from list_chats over a raw handle. Messages refuses to enumerate participants for a script, so an existing conversation is the only target that can be addressed reliably — but chat carries an id, and the store holds a guid for every conversation. The read lane picks the target and the write lane addresses it by id, which is the only arrangement where neither lane has to do the thing it cannot.

Apple Events then hands back nothing at all — no identifier for what it sent. So the client takes a timestamp before the send and polls chat.db for the outgoing row:

`pending` is not a failure. Do not retry it — the message was sent, and a retry sends it twice.

No send has yet been executed against a live Messages. The lane is written from the dictionary and from the store, and its tests run against a fake osascript on purpose. docs/messages.md records exactly what that leaves open; the safe way to measure it is a message to your own handle.

Configuration

APPLE_MESSAGES_ALLOW_CODES gates find_codes, which extracts one-time 2FA codes from recently received messages. It is a read, and it is deliberately not folded into ALLOW_WRITES: reaching a read through the write gate would mean granting the right to send a message in order to get it.

It is gated at all because of what it combines with. This server already holds the conversation history and @mgcrea/mcp-apple-mail holds the inbox — between them, the password-_reset_ channel. Adding live authentication codes completes an account-takeover primitive out of parts that were each individually reasonable, so it defaults off and the two gates are independent in both directions.

Codes are matched by signal rather than by a \d{4,8} regex, and every result carries a confidence and a matched saying how it was found; anything below high should be checked against the message body before use. See src/codes.ts in @mgcrea/mcp-apple-core — the heuristic moved there when @mgcrea/mcp-apple-safari became its second caller — and docs/passwords.md in the repo for why the Passwords app itself is unreachable and this is what ships in its place.

Notes that will bite you

• Dates do not fit in a JavaScript number. Every date column is nanoseconds since 2001 — eighteen digits, past Number.MAX_SAFE_INTEGER — and node:sqlite _throws_ rather than truncating. Swallowed by a try/catch that throw looks exactly like "this column is empty", and it is how a probe once reported "no dates present" for all seven columns across 97,414 messages. Every query here divides in SQL so the integer never reaches JavaScript.
• Tapbacks are rows in the message table. 2,788 of them on the measured store, and a reader that does not filter them renders Liked "see you at 8" as if somebody typed it. They are excluded from conversations by default and reported on the message they target.
• Search covers the blob-only messages. A LIKE pass over the column (16 ms across 97,416 rows) plus a decode pass over the 3,051 rows SQL cannot see (~6 ms). Completeness turned out to be nearly free; a search that silently omitted 3% would not have been.
• Refs are GUIDs, not rowids. SQLite reuses a deleted row's id, and Messages deletes constantly, so a rowid handed out in one turn can resolve to a different message two turns later — plausible, wrong, and silent.
• A send cannot be taken back, and nothing here pretends otherwise. There is no unsend, no draft and no preview in the dictionary; send delivers. The tool is gated on ALLOW_WRITES, requires `confirm: true