Fleetpost

A mailbox for your machines. Independent AI coding agents on separate computers drop tasks in a shared folder and pick them up whenever they wake — no server, no daemon, and no requirement that both sides be online at once.

Offline-tolerant coordination for independent AI agents: cross-machine, server-less, carried by any folder rclone can reach — Google Drive, S3, Dropbox, SFTP, WebDAV.

Fleetpost is a handful of shell scripts and a convention. You run several agents (Claude Code, Cursor, Codex, Aider, whatever) on several machines. Each keeps its own local memory. Fleetpost lets them know what the others can do and ask each other for work — and it keeps working when a machine is asleep or powered off, because the messages simply wait in the shared folder until that machine's next sync.

Why this exists

Run an agent on your laptop and another on your desktop and they are blind to each other; you become the courier, copy-pasting between terminals (a real, common pain). The usual answers are live — both agents must be running and connected at the same moment. Fleetpost takes the other route: asynchronous. Leave the work in the drop and walk away; the powered-off machine catches up on its own.

How it works

<shared-folder>/                     (any rclone remote: Drive, S3, Dropbox…)
├── 00-START-HERE.md                 onboarding for a new machine
├── PROTOCOL-CHANGES.md              append-only changelog, watched by hash
├── laptop/
│   ├── capabilities.md              what this machine can do (hand-written)
│   ├── inventory.md                 what it knows (index; optional generator)
│   ├── last-sync.txt                when it last ran a cycle (its heartbeat)
│   └── inbox/
│       └── handled/                 requests it has completed
├── desktop/
└── server/

A small sync.sh runs on a timer (systemd / cron / launchd) on each machine and does four things — and nothing that touches another machine's data:

1. pulls this machine's inbox/
2. detects new requests (and protocol changes) → raises a local SIGNAL.md flag
3. pulls every other machine's capabilities.md and last-sync.txt, and looks at their inboxes to see which of this machine's own requests are still unhandled
4. publishes this machine's own descriptors (only when changed) and its heartbeat

It never executes a request. It fetches and flags; the agent does the work in a session.

Demo

Your desktop needs something only the laptop can do, so it drops a request in the laptop's inbox and forgets about it:

desktop$ echo "Compile the arm64 build and report the sha256." \
           > 2026-08-09-from-desktop-compile.md
desktop$ rclone copy 2026-08-09-from-desktop-compile.md \
           "gdrive:coordination/laptop/inbox/"

The laptop is asleep. Hours later it wakes; its timer fires a sync:

laptop$ ./scripts/sync.sh
14:17  1/4 pulling laptop/inbox …
14:17  2/4 detecting new requests and protocol changes …
       -> new requests; raised SIGNAL.md
14:17  3/4 pulling fleet capabilities …
       -> desktop: fetched
14:17  4/4 publishing my descriptors (only if changed) …
14:17  done.
# exit code 10 = "something new for you"

laptop$ cat ~/.agent-coordination/SIGNAL.md
# SIGNAL — something is waiting for you
## New requests in your inbox (~/.agent-coordination/inbox/)
- `2026-08-09-from-desktop-compile.md` (41 bytes)

The laptop's agent does the work, drops a reply in desktop/inbox/, and moves the request into inbox/handled/. The desktop picks up the answer on its next sync — and at no point did both machines need to be online at the same time.

What makes it different

The category is not empty — see Alternatives. Fleetpost's specific bundle is:

• rclone-first transport — pull from object storage / Drive / S3 / Dropbox. No sync daemon on every host, no git remote. If rclone can reach it, Fleetpost can use it.
• Folder-native protocol — per-machine inbox/ + handled/ subfolders. A machine with no automation still sees exactly its unhandled requests at the top level.
• Offline-published capability descriptor — each machine leaves a "what I can do" file in the folder, discoverable even while that machine is off (unlike a live Agent Card that needs an HTTP server up).
• Hash-watched changelog — rule changes reach already-running agents, which would otherwise never re-read the onboarding doc.

Use it from an agent (MCP)

There is an MCP server in mcp/ — fleetpost-mcp on PyPI — so an agent can read the fleet's capabilities, work through its own inbox, and hand a task to another machine as tool calls:

{ "mcpServers": { "fleetpost": {
    "command": "uvx", "args": ["fleetpost-mcp"],
    "env": { "FLEETPOST_CONFIG": "/path/to/fleetpost/config.env" } } } }

It reads this same config.env and runs these same scripts — no second implementation of the protocol, and still no server. See mcp/README.md.

Quickstart

git clone https://github.com/StanimirTenev/fleetpost
cd fleetpost
./scripts/init.sh             # asks four questions, writes config.env, claims your folder
$EDITOR ~/.agent-coordination/self/capabilities.md   # what this machine can do — the fleet reads it
./scripts/doctor.sh           # confirm the wiring before wondering why nothing arrives
./scripts/sync.sh             # first cycle: publish yourself, fetch the others

# one-time for the whole fleet: seed the shared docs (from any one machine)
rclone copy templates/00-START-HERE.md      "<remote>:<coord-dir>/"
rclone copy templates/PROTOCOL-CHANGES.md   "<remote>:<coord-dir>/"

# schedule it (Linux, user timer)
cp examples/systemd/coordination-sync.* ~/.config/systemd/user/
systemctl --user daemon-reload
systemctl --user enable --now coordination-sync.timer
loginctl enable-linger "$USER"

Prefer to wire it by hand? cp config.example.env config.env and edit — init.sh writes exactly that file and nothing else.

Day to day

./scripts/status.sh    # what's waiting for me, what I'm still waiting on, what the fleet
                       # can do (local, no network)
./scripts/send.sh --to desktop --topic "sign the installer" \
                  --want "Sign dist/app.exe with the company cert." \
                  --done "signtool verify /pa passes on the uploaded file." \
                  --until 2026-09-01
./scripts/handle.sh <filename>   # done with one: move it to inbox/handled/

send.sh requires every field the protocol asks for, so a request that cannot be acted on cannot be created.

Notice requests without being told

An established agent never re-reads the onboarding doc, so a request can sit unnoticed. examples/hooks/session-start.sh prints whatever the last cycle flagged at the top of an agent session, and stays completely silent when nothing is waiting. It reads local files only — set FLEETPOST_HOOK_SYNC=1 to pull first, which is the right choice on a machine with no scheduler.

See docs/PROTOCOL.md for the full protocol and 00-START-HERE.md for what a new agent reads.

Requirements

bash, rclone (configured with one remote), sha256sum, and a scheduler (systemd, cron, or launchd). That's it. No server, no database, no language runtime.

Alternatives

Honest comparison — pick what fits:

If you already run Syncthing everywhere, SAMP is excellent and closest in spirit. Fleetpost is for when your shared layer is object storage or a cloud drive (via rclone) and you want folder-native inbox/handled semantics plus offline capability discovery.

Design principles

• Memory stays local. Only a short descriptor is published. Full memory never travels.
• Ask for actions, not access. Secrets never travel; the machine that holds a credential does the work and returns only the result.
• One author per shared document, so concurrent writers can't clobber a changelog entry.

License

MIT.