agentleFS
TechNickAITechNickAI/hermes-skills/skills/productivity/imessage-bluebubbles/SKILL.md

imessage-bluebubbles

Skill0 starsChanged 3 months agoMITView on GitHub

Use when sending, reading, or searching iMessages from an agent on macOS, or when setting up, hardening, or debugging the BlueBubbles iMessage bridge. Also use when an existing chat.db/imsg approach breaks with permissionDenied, "authorization denied (code 23)", or hangs with no output after a macOS or Python upgrade.

  • Reads credentials
  • Installs packages

The file

We split this skill into the rules it gives an agent

12 rules. About 0 are common, 8 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈8

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use for ANY iMessage work on a Hermes fleet Mac -- sending, reading, searching messages, or standing up the bridge itself.

    line 25
  2. Distinctive:

    Read the TCC table to explain _why_ something is broken.

    line 133
  3. Distinctive:

    Never talk a non-technical owner through disabling SIP for typing indicators -- not worth it, and on Apple Silicon it also disables running …

    line 172
  4. Very specific:

    Treat a public server_address on any machine as a finding to fix, not a note to file.

    line 232
  5. Distinctive:

    Never message an unknown number without explicit approval.

    line 258
  6. Very specific:

    Run health before diagnosing anything.

    line 260
  7. Very specific:

    Run the suite. It exercises the bridge the way an agent actually uses it and reports PASS/FAIL/SKIP per check.

    line 280
  8. Distinctive:

    Auth enforcement is tested with a WRONG password: If a bad password is accepted, every other PASS is meaningless.

    line 296
  9. Distinctive:

    Verify in three layers, because each proves something the others do not:

    line 320
  10. Distinctive:

    When counting name duplicates, separate same-root duplicates (real, breaks the loader) from local-shadows-bundled pairs (normal -- the …

    line 330
  11. Distinctive:

    Check the length, not just presence.

    line 391
  12. Very specific:

    Run both in parallel for a few days before retiring imsg.

    line 424

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

iMessage via BlueBubbles

Use for ANY iMessage work on a Hermes fleet Mac -- sending, reading, searching messages, or standing up the bridge itself. Also the right skill when iMessage is broken: imsg failures (permissionDenied, authorization denied (code 23), hangs with no output) are the signal to move to this bridge.

Why this exists

The older imsg CLI reads Apple's ~/Library/Messages/chat.db directly. macOS governs that file with TCC, and TCC grants Full Disk Access to an exact interpreter path. So a routine uv Python patch bump silently revokes access and every scheduled iMessage job dies with authorization denied (code 23) -- no Hermes version change, no config change, no log warning. That happened on a-contact a real incident. Add the sandboxd wedge (imsg hangs forever, zero output) and direct-database access is structurally fragile. BlueBubbles moves the Apple-facing half into one long-running signed app that holds its own permission grant. Hermes talks to it over loopback HTTP. Python never touches chat.db again, so a Python upgrade can no longer break iMessage. That is the whole point -- a structural fix, not a workaround. Hermes ships a first-class BlueBubbles platform adapter (gateway/platforms/bluebubbles.py, registered as Platform.BLUEBUBBLES), so the agent can receive iMessages as a real channel, same as Telegram.

Two separate capabilities -- do not confuse them

CapabilityMechanismUse for
Inbound channelHermes gateway adapter, BLUEBUBBLES_* env, webhook on:8645People messaging the agent and it replying in-thread
Outbound toolscripts/bb.py calling the REST APIThe agent sending or reading messages as part of a task

Setting the env vars enables the channel. bb.py works independently and needs only the same two credentials.

Setup

./scripts/setup-bluebubbles.sh              # full guided install
./scripts/setup-bluebubbles.sh --verify     # health-check an existing install
./scripts/setup-bluebubbles.sh --exposure   # check for public tunnel exposure
./scripts/setup-bluebubbles.sh --wire # write BLUEBUBBLES_* to.env only

What cannot be automated, ever

macOS protects the TCC database with SIP. It can be read with sudo for diagnosis, but it cannot be written -- there is no supported way to grant a permission from a script. Permission grants are a human click. Do not claim setup is done until a real API call proves access. Budget a screenshare for a remote machine. The human steps are exactly three:

  1. Set a server password in BlueBubbles' first-run wizard.
  2. Grant Full Disk Access to BlueBubbles when macOS prompts.
  3. Turn the public tunnel off (see Hardening). On Automation > Messages: BlueBubbles may also prompt for Automation access to Messages. Grant it if asked. Do not preemptively send someone to System Settings for it, and do not diagnose a slow send as a missing Automation grant: sending has been verified working with no Automation TCC entry present, because an app can inherit the right from an authorized parent process. Full Disk Access governs reading chat.db; Automation governs driving Messages.app. They are different permissions, but only the first is reliably required up front. See references/send-path-diagnosis.md before acting on a hang.

Skip the Google / Firebase sign-in -- Hermes does not use it

First-run setup pushes a Google sign-in. That is Firebase Cloud Messaging, which exists so the BlueBubbles phone app receives push notifications. The Hermes adapter receives messages by webhook over loopback (127.0.0.1:8645), so FCM is never consulted. Skip it. If a fleet owner gets stuck on Google auth, that is not a blocker -- it is an optional step for a client we do not run. The same reasoning applies to the Cloudflare tunnel (see Hardening): both exist for remote phone clients. the operator hit exactly this on a real incident and it stalled the install for no reason.

Full Disk Access will NOT be in the list until the app asks for it

macOS populates the Full Disk Access list on demand -- an app appears only after it first attempts to read a protected file. A half-configured BlueBubbles has not tried to open chat.db yet, so it is genuinely absent from System Settings, and the user has not missed anything. The order is therefore the reverse of what the setup instructions imply:

  1. Finish first-run setup (server password set)
  2. BlueBubbles attempts to read chat.db
  3. Then it appears under Full Disk Access
  4. Enable it, then restart BlueBubbles -- TCC is evaluated at process start, so a running app never picks up a new grant Do not send someone hunting through Settings for an entry that cannot exist yet. Confirm what is actually granted by reading TCC directly:
sudo sqlite3 "/Library/Application Support/com.apple.TCC/TCC.db" \
  "select client, service, auth_value from access where client like '%luebubble%';"

auth_value 2 = granted, 0 = denied, no row = never requested. kTCCServiceAccessibility is the Private API grant (tapbacks, typing); kTCCServiceSystemPolicyAllFiles is Full Disk Access. They are separate -- having Accessibility does not imply FDA. But do not treat a missing FDA row as proof of no access. Verified on the operator's Mac a real incident: BlueBubbles returned the full chat list with no kTCCServiceSystemPolicyAllFiles entry at all, only Accessibility. An Electron app launched from an already-authorized parent can inherit disk access, so the TCC table under-reports. The authoritative test is always a real chat/query call:

./scripts/bb.py health      # 'chat access: ok' is the only proof that counts

Read the TCC table to explain why something is broken; never to conclude that something works or does not.

Gatekeeper: the app refuses to launch until you clear quarantine

BlueBubbles is signed with a Developer ID but not notarized by Apple. macOS tags every download with com.apple.quarantine, and on an un-notarized app that tag makes it refuse to open (often as a misleading "damaged" message). The fix is one command, and setup-bluebubbles.sh now does it automatically:

xattr -dr com.apple.quarantine /Applications/BlueBubbles.app

This clears the download tag on one bundle only. It does not weaken Gatekeeper system-wide. Verified in practice -- the app launched immediately afterward. Two follow-on facts:

  • spctl -a /Applications/BlueBubbles.app still reports rejected / Unnotarized Developer ID after the tag is cleared. Expected, and does not block launch. Do not chase it.
  • GUI fallback: System Settings > Privacy & Security, scroll down, Open Anyway.

The Homebrew cask is deprecated -- prefer the DMG for fleet rollout

brew install --cask bluebubbles works today but is deprecated for failing the Gatekeeper check and is disabled one occasion. Convenient because it is scriptable, but not the durable path. For fleet rollout use the DMG from https://bluebubbles.app/downloads/ plus the quarantine-clear above.

SIP is optional -- default to leaving it ON

SIPWhat worksWho
Enabled (default)Send, receive, attachments, read receiptsNon-technical owners. Recommended.
DisabledAdds tapbacks, typing indicators, edit/unsendTechnical hosts already running SIP-off

Sending and receiving do not need SIP disabled. Only the Private API extras do. Never talk a non-technical owner through disabling SIP for typing indicators -- not worth it, and on Apple Silicon it also disables running iOS apps. Prior fleet notes recording "BlueBubbles needs SIP disabled + ~2hr of pain" describe the Private API tier, not the base install.

Hardening: turn off the public tunnel

BlueBubbles opens a public Cloudflare tunnel by default. On first launch the a first launch generates a live trycloudflare.com URL pointing at the host. That exists so phone clients can reach the server from anywhere -- but Hermes talks to it over loopback and never needs it. Left on, it is a public URL fronting the owner's entire message history, guarded by one password. Turn it off: BlueBubbles > Settings > Connection, set Proxy Service to Dynamic DNS with address http://localhost:1234, or disable the tunnel.

Turning it off headlessly, and the orphaned-process trap

Config is held in memory and flushed on shutdown (set-config / config-update handlers in app.asar), so a direct DB write while the app runs gets clobbered. Correct sequence:

# 1. quit cleanly so in-memory config is flushed
osascript -e 'tell application "BlueBubbles" to quit'
# 2. write the DB (values verified from the app bundle: dynamic-dns | lan-url |
#    cloudflare | ngrok)
sqlite3 ~/Library/Application\ Support/bluebubbles-server/config.db \
  "update config set value='dynamic-dns' where name='proxy_service';
   update config set value='http://localhost:1234' where name='server_address';"
# 3. relaunch and verify the values SURVIVED the restart
open -a BlueBubbles

Then kill the orphaned tunnel. Verified in practice: after the config change and a full app restart, the bundled cloudflared was still running from the previous session and the public URL still answered HTTP 200. The config was correct and the box was still exposed.

pgrep -f "BlueBubbles.app.*cloudflared"     # BlueBubbles' own tunnel only
kill <pid>

Match on the BlueBubbles.app path, never on cloudflared alone -- the operator may run their own unrelated tunnels on the same box. Killing those is collateral damage. Proof is a dead public URL, not a changed setting:

curl -m 25 -o /dev/null -w "%{http_code}\n" "https://<old-url>/api/v1/ping"
# 530 or connection failure = actually closed. 200 = still exposed.
./scripts/setup-bluebubbles.sh --exposure   # reads the server's own config

Treat a public server_address on any machine as a finding to fix, not a note to file.

Daily use

scripts/bb.py is the agent-facing replacement for imsg. Plain-text output, one record per line -- no JSON parsing.

./scripts/bb.py health                                   # server + auth + disk access
./scripts/bb.py chats --limit 20                         # recent conversations
./scripts/bb.py find --query "alex"                  # locate a chat
./scripts/bb.py history --chat "alex" --limit 30     # read a conversation
./scripts/bb.py send --chat "alex" --text "on my way"

--chat accepts a raw GUID or a fuzzy name/number. On an ambiguous match it prints the candidates and exits without sending -- messaging the wrong person is not recoverable, so it never guesses.

Rules

  1. Confirm recipient and content before sending on the user's behalf.
  2. Read the thread before sending to a recipient you resolved yourself. Stored contact data conflicts and Apple Contacts genuinely attaches one number to multiple people. See references/recipient-verification.md.
  3. Never message an unknown number without explicit approval.
  4. Rate-limit yourself. No bulk sending.
  5. Run health before diagnosing anything.
  6. Never retry a send after a timeout. Re-read the thread to establish whether it landed -- a timeout is absence of information, not failure.

Verification

ping is not sufficient -- it passes with no disk access at all. The real test is chat/query, which reads the message database:

./scripts/bb.py health     # exercises ping, server/info, AND chat/query

chat access: ok is the only line that proves Full Disk Access works. A green ping with empty chats means the app is running but blind. health exits non-zero in that case, so a caller can gate on the status code.

Full test suite

Run the suite. It exercises the bridge the way an agent actually uses it and reports PASS/FAIL/SKIP per check.

./scripts/test-bluebubbles.py                      # read-only, safe, no messages
./scripts/test-bluebubbles.py --send-to "any;-;+15551234567" \
                              --from-name "YourAgent"
./scripts/test-bluebubbles.py --json               # machine-readable summary

Exit 0 = all passed, 1 = failures, 2 = could not run. 16 checks across seven areas: process, auth, disk access, read path, safety guards, network exposure, Hermes wiring, and (opt-in) a live send. Design rules baked into the suite, each earned from a real failure:

  • Auth enforcement is tested with a WRONG password. If a bad password is accepted, every other PASS is meaningless.
  • Disk access is tested with chat/query, never ping. Ping succeeds with zero disk access.
  • Exposure is tested by fetching the public URL, not by reading a config value. Config can say loopback while an orphaned tunnel still serves traffic.
  • A live send is confirmed by reading the message back from the server, not by trusting the sender's own success claim.
  • The send test is opt-in and requires an explicit GUID, because it texts a real human. Identify yourself in test messages. Validate the suite itself periodically by pointing it at a wrong password and a wrong port; both must FAIL. A suite that cannot fail proves nothing. bb.py health remains the quick single check; chat access: ok is the only line that proves Full Disk Access.

Fleet rollout

This skill is macOS-only -- BlueBubbles is a Mac app. Delivering it to a Linux host produces a skill that can never work, which is worse than absent. Resolve the OS per host before copying (ssh <host> 'uname -s', Darwin only).

Deliver a tarball rather than per-file copies, extract to a staging directory, and only swap the live directory once SKILL.md is confirmed present -- a partial transfer must never half-replace a working skill. Back up any existing copy first.

Verify in three layers, because each proves something the others do not:

  1. Delivery -- sha256 every file against the source. File-present is not file-correct.
  2. Loader -- resolve the skill BY NAME on the target and confirm exactly one definition. A duplicate name makes skill_view refuse, so a perfectly delivered skill can still be unloadable.
  3. Body -- assert a distinctive string survived (chat access: ok, UNKNOWN -- DO NOT RETRY).

When counting name duplicates, separate same-root duplicates (real, breaks the loader) from local-shadows-bundled pairs (normal -- the local copy wins). Any machine whose local skill library overlaps the bundled one shows many of the latter, and reporting them as collisions is a false alarm. Only a same-root duplicate is a defect.

Profile layout trap: on a single-agent host the agent IS the root profile and skills live at ~/.hermes/skills/, with no profiles/<name>/ directory. Sub-profiles use ~/.hermes/profiles/<name>/skills/.

Pitfalls

A send that times out has NOT necessarily failed -- verify, never retry

The most dangerous failure mode in this skill. AppleScript sends routinely exceed a 20-30s HTTP timeout: the request blocks while Messages.app does the work, but the message still sends. The client gives up; Apple does not. Seen in practice." It had sent. She replied. the operator caught the error. Two rules follow:

  1. A timeout means UNKNOWN, not failed. Absence of information is not evidence of absence. Never report a send as failed on a timeout alone.
  2. Verify after a delay, and never retry blind. An immediate re-read races the in-flight send and returns a false negative. Wait, re-read the thread, match on message text. Retrying instead of verifying double-texts a real person -- unrecoverable. bb.py send handles this: 120s timeout, then polls the thread for 30s looking for the sent text, printing CONFIRMED delivered or an explicit UNKNOWN -- DO NOT RETRY that tells the operator to check Messages.app before retrying. By hand, the same discipline:
# WRONG -- races the send, reports a false negative
./bb.py send --chat X --text "hi";./bb.py history --chat X --limit 3
# RIGHT -- give a slow send time to land before concluding anything
./bb.py send --chat X --text "hi"; sleep 30;./bb.py history --chat X --limit 3

A hang is not proof of a missing permission

A hanging send looks exactly like a blocked Automation prompt, so it is tempting to conclude the app lacks Automation access to Messages (com.apple.MobileSMS). On the In testing that inference proved wrong -- sends worked with no such TCC grant; the hang was pure latency, and the user was nearly sent to System Settings for nothing. Distinguish by error shape: a missing permission or helper returns an explicit error (iMessage Private API Helper is not connected!), while latency hangs silently. Reads working tells you nothing about sends -- they use different mechanisms -- but neither does a hang tell you a permission is missing.

Name search returns nothing for someone you know exists

iMessage chats are keyed by phone number and email; displayName is null on most 1:1 threads. A name finding nothing is expected and is not evidence the person has no thread. Search by number fragment, and paginate -- chat/query caps at 1000 rows per call.

ping succeeds but no chats come back

Full Disk Access is not granted. Grant FDA to BlueBubbles.app, then restart the app -- TCC is evaluated at process start, so a running process never picks up a new grant.

Password shows as set but auth still fails

Check the length, not just presence. First launch can leave an empty password that still appears in config:

sqlite3 ~/Library/Application\ Support/bluebubbles-server/config.db \
  "select name, length(value) from config where name='password';"

Length 0 means first-run was never completed. Finish the wizard.

open -a BlueBubbles hangs forever

The Mac is at the login screen. GUI apps cannot launch while locked, and the call blocks rather than failing. Take a screenshot (screencapture -x) before assuming the app is broken. Remote fleet Macs must be unlocked and logged in.

Port 1234 already taken

Set BB_PORT and re-run. Preflight distinguishes "held by BlueBubbles" (fine) from "held by something else" (fatal).

Verifying over SSH proves nothing about scheduled jobs

Inherited from the imsg era, still applies to anything TCC-related: over SSH the responsible process is sshd, which already holds Full Disk Access, so a probe can pass while every launchd cron job is denied. Only output from a job that fired on its own schedule is authoritative. BlueBubbles largely sidesteps this -- the grant belongs to the always-running app, not to whatever interpreter invokes it -- which is precisely why it is more robust than imsg.

Migrating off imsg

Run both in parallel for a few days before retiring imsg. Keep the imessage skill installed during the soak: it documents the sandboxd watchdog and the TCC diagnostic ladder, still the right tools for a genuinely wedged Messages stack. Retire imsg only after bb.py health passes from a real scheduled run, not from an interactive shell.

References

  • references/api-surface.md -- BlueBubbles REST endpoints Hermes uses
  • references/landscape-2026.md -- why BlueBubbles over Beeper, pypush, or a paid API
  • references/send-path-diagnosis.md -- sends hang or 500 while reads work; Automation grant, Private API Helper, GUID format, pagination
  • references/recipient-verification.md -- confirming you have the right person before an irreversible send

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.