Reported issues for Apple Mail MCP by imdinu

Pod holds 20 of 77 GitHub reports that passed its relevance review. This can include external user reports, maintainer-confirmed bugs, and concrete feature gaps. Treat them as evidence to inspect, not a count of distinct defects.

Back to Apple Mail MCP by imdinu.

Most discussed

get_email type coercion + search mailbox scope + scheduled mode failures

Bug Report — Three related issues in scheduled/automated mode

Environment

• Version: v0.1.8
• Running via: Claude Cowork scheduled tasks (automated, unattended)

Bug 1: get_email — message_id type coercion in scheduled mode

When called from a scheduled task, message_id is passed as a string instead of an integer. The tool rejects it with a type error, making it impossible to read email bodies in any automated context.

Workaround: None reliable. Triage falls back to m

Read the thread · 2026-03-27 · closed · external user · 3 comments

Large mailboxes timeout on startup

I have a lot of emails (400,000+). I can build the index, but every time I open Claude, the MCP server takes too long to traverse all the email (needs about 2 minutes) and as a result Claude pops up an error message saying that it cannot connect to the MCP server.

It would be great if you can find a way to support large mailboxes. Perhaps the software could do the start-up scan in the background, so that it can still respond to Claude when it tries to connect to the MCP server?

Thanks

Read the thread · 2026-03-04 · closed · external user · 3 comments

search() cannot distinguish a missing index from a genuine no-match, so an unbuilt index reads as an empty mailbox

search() cannot distinguish "no index" from "no match", so an unbuilt or stale index looks exactly like a genuinely empty result. On my install that cost roughly four months of silently empty searches, on a mailbox of 18,232 emails that would have matched constantly.

Filing this with the backing of the Dex project, where the same problem is tracked as davekilleen/Dex#446. Dave (the maintainer there) asked me to bring it here directly since I have the reproduction. Happy to send a PR if the sh

Read the thread · 2026-08-12 · closed · outside contributor · 2 comments

PermissionError escapes both no-FDA fallbacks: get_emails() fails outright, and attachment reads report "not found"

Running without Full Disk Access — which the docs describe as supported ("The MCP server itself does not need Full Disk Access — only the index and rebuild commands do", docs/troubleshooting.md) — two code paths fail in ways that look like different bugs but share one root cause.

PermissionError subclasses OSError, not FileNotFoundError. Both sites below have a handler that clearly intends to cover an unreadable Mail store, and PermissionError walks straight past it.

Happy to

Read the thread · 2026-08-07 · open · external user · 2 comments

Feature request: APPLE_MAIL_INDEX_EXCLUDE_ACCOUNTS for account-level index exclusion

Summary

Request a new environment variable APPLE_MAIL_INDEX_EXCLUDE_ACCOUNTS (comma-separated, like the existing APPLE_MAIL_INDEX_EXCLUDE_MAILBOXES) that excludes entire accounts from the index at build time.

Motivation

APPLE_MAIL_INDEX_EXCLUDE_MAILBOXES matches by mailbox name only, with no account-qualified syntax. This works fine for excluding category folders like Drafts (the default), but fails for the common case of wanting to exclude an entire account whose mailboxes have g

Read the thread · 2026-05-22 · closed · external user · 2 comments

Attachment Download Error

Always got error: attachment not found when run with cloude coworker. But can got it when via MCP Inspector.

get_email_attachment Request

{ "message_id": 437629, "filename": "GCEMEARSD0000290714-2.pdf", "account": "info@playlingo.app", "mailbox": "[Gmail]" } Error

Error calling tool 'get_email_attachment': Attachment 'GCEMEARSD0000290714-2.pdf' not found in email 437629.

Read the thread · 2026-04-03 · closed · external user · 2 comments

Use SQL-based diffing in sync instead of in-memory inventory comparison

Problem

sync_from_disk() loads the full disk inventory and full DB inventory into Python dicts to compute the diff. For users with 200K+ emails, this causes significant memory spikes.

Proposed Solution

1. Create a temporary SQLite table (disk_inventory_temp)
2. Walk the disk, inserting paths into the temp table
3. Use SQL EXCEPT or LEFT JOIN queries to find NEW/DELETED/MOVED
4. Let SQLite handle the diffing — optimized for this and uses disk, not RAM

Current Impact

Low for

Read the thread · 2026-03-23 · closed · 2 comments

Two concurrent --watch instances (Claude Desktop spawns MCP servers twice) cause write storm, unbounded WAL growth and blocked FTS searches

Environment

• apple-mail-mcp 0.4.2
• macOS 26.5.2
• Claude Desktop as MCP client
• Observed on 2026-07-15

Summary

Claude Desktop structurally spawns two instances of every configured MCP server — one for Desktop chats and one for the Claude Code bridge. When both instances run with serve --watch, the two watchfiles watchers operate on the same SQLite database (~/.apple-mail-mcp/index.db, WAL mode) and get into a sustained write storm.

Symptoms

• Log spam every few seconds: `In

Read the thread · 2026-07-15 · closed · external user · 1 comment

Most recent

get_emails() silently returns [] for Gmail-backed accounts: Envelope Index fast path misses label-based mailbox membership

Problem

On 0.4.0/0.4.1, get_emails(account, mailbox) returns an empty list for Gmail-backed accounts even when the mailbox has plenty of mail. No error is raised, so the JXA fallback never runs and the caller gets a silently wrong empty result.

Repro (real mailbox, macOS 14, Mail V10)

get_emails(account="<gmail account>", mailbox="INBOX") returned [] for an INBOX that AppleScript confirms holds 104 messages.

Root cause

The Strategy-0 fast path assumes mailbox membership is ex

Read the thread · 2026-06-12 · closed · outside contributor · 0 comments

search() docstring: document the empty-result hint shape

Problem

search() declares -> list[SearchResult] | dict and returns {"result": [], "hint": "..."} when nothing matches, but the docstring's Returns section only describes the list shape. The hint-on-empty pattern is good LLM ergonomics and should stay — it just needs to be documented so clients and harnesses iterating the result know both shapes exist.

Proposed Solution

Document the empty-result shape in the Returns section. (The larger return-type unification question is part of t

Read the thread · 2026-06-11 · closed · 0 comments

MCP boundary: limit/offset unclamped, before/after dates unvalidated, STRATEGY3_* env vars unbounded

Problem

Three input-validation gaps at the MCP tool boundary, all producing silent misbehavior rather than correctable errors:

1. limit/offset are unclamped. Negative values flow into SQL, where LIMIT -1 means unlimited in SQLite. Huge limits force the full result set into memory and into the model's context. They're also inlined into JXA scripts via builders.py.
2. before/after aren't validated. The docstring promises YYYY-MM-DD, but a malformed date like 2026-6-1 g

Read the thread · 2026-06-11 · closed · 1 comment

Watcher: no rollback on sqlite3.Error; a failed batch can poison the next commit

Problem

watcher.py::_process_pending() runs deletes and adds inside one try, commits once at the end, and the outer except sqlite3.Error only logs. Python's sqlite3 opens an implicit transaction on the first DML and holds it until commit/rollback — so a mid-batch failure leaves the connection sitting in an open transaction with partial work. The next batch's commit() then persists that stale partial state alongside the new batch.

Related: both watcher.py and sync.py fetch ro

Read the thread · 2026-06-11 · closed · 1 comment

Inline images without MIME filename are dropped from _extract_attachments

Problem

_extract_attachments in disk.py skips MIME parts that have no filename attribute and aren't marked Content-Disposition: attachment, even when they have a Content-ID header and are physically stored on disk by Apple Mail.

Real example from a multipart/related HTML email in a 72K-message mailbox:

Part [4]:
  Content-Type: image/png
  Content-Disposition: inline
  Content-ID: <RO_BCR_Email_94c8b886-d7d1-4c...>
  filename: <none>

Apple Mail saved the image to `Atta

Read the thread · 2026-05-14 · closed · 0 comments

Sync inventory walk dominates wall-clock at >100k mailboxes — disk-walk + per-file parse layer

Problem

The SQL-temp-table sync diff (#60) shipped in v0.3.0 and works as designed at the diff layer. But for users with very large mailboxes (>100k messages), wall-clock time for apple-mail-mcp index is dominated by the layer before the diff — the unconditional disk inventory walk plus per-file parse — not the diff itself.

Source

External report from @chrispyness77 in #60 (comment).

Three back-to-back `a

Read the thread · 2026-05-10 · open · 0 comments

Read Apple's Envelope Index directly for metadata ops (list_accounts, list_mailboxes, get_emails, search_subject)

Background

The v0.3.0 benchmark refresh (~72K-message mailbox) revealed that two competitors — BastianZim/apple-mail-mcp and like-a-freedom/rusty_apple_mail_mcp — are dramatically faster than us on every metadata-only operation:

Read the thread · 2026-05-07 · closed · 1 comment

Memory spike in _estimate_attachment_size from chained .replace() calls

Problem

disk.py::_estimate_attachment_size decodes the encoded payload via part.get_payload(decode=False), then computes the decoded byte size by stripping whitespace and counting:

raw = part.get_payload(decode=False)
# ...
stripped = raw.replace("\n", "").replace("\r", "")
clean_len = len(stripped.replace(" ", ""))

For a 20 MB base64 attachment (well under the 25 MB MAX_EMLX_SIZE cap), this allocates three intermediate string copies of approximately 20 MB each — ~8

Read the thread · 2026-05-07 · closed · 0 comments

APPLE_MAIL_READ_ONLY config flag is set up but never enforced at MCP tool boundaries

Problem

config.py exposes APPLE_MAIL_READ_ONLY (env var) and set_read_only_mode() / is_read_only() accessors. cli.py::serve accepts --read-only and calls set_read_only_mode(True). The flag is documented in CLAUDE.md and README.md as a way to "Disable write operations".

But: there are no write operations in v0.3.0. Every MCP tool currently registered (list_accounts, list_mailboxes, get_emails, get_email, search, get_email_links, get_email_attachment, `get_atta

Read the thread · 2026-05-07 · closed · 0 comments

Attachment cache files inherit umask permissions (should be 0o600)

Problem

server.py::_cleanup_old_attachments and the attachment-extraction path in get_email_attachment write extracted attachment content under ~/.apple-mail-mcp/attachments/. The cache directory is created with mode 0o700 (good — only the owner can list it). But the files themselves are written via Path.write_bytes() (or whatever the extraction path uses), which means they inherit the user's umask — typically 0o644 on macOS.

On a single-user machine this is fine. On a share

Read the thread · 2026-05-07 · closed · 0 comments

index://status resource walks the disk on every call (need TTL cache for disk_email_count)

Problem

The index://status MCP resource (added in v0.3.0 via #12) calls IndexManager.get_stats(), which calls get_disk_inventory(mail_dir) to populate disk_email_count. That walks the filesystem under ~/Library/Mail/V*/ — potentially 100K+ files for users with large mailboxes.

Today the call is wrapped in asyncio.to_thread() to avoid blocking the event loop, but if an MCP client polls the resource on a tight loop (some agent UIs do), the asyncio thread pool gets pinned and the d

Read the thread · 2026-05-07 · closed · 0 comments

Watcher silently swallows DLQ insert failures

Problem

In index/watcher.py::_process_pending, when the failed-parse DLQ insert (added in v0.3.0 via #58) raises sqlite3.Error, the failure is logged at WARNING level but the underlying error is otherwise swallowed:

except sqlite3.Error as e:
    logger.warning(
        "Could not record parse failure for %s: %s",
        path, e,
    )

This is fine for a one-off transient error (lock contention, etc.). But if the disk fills up, the DB is corrupted, or the schema is someh

Read the thread · 2026-05-07 · closed · 0 comments

The remaining reports are on the project's issue tracker.