Swarmail

swarmail lets coding-agent sessions send each other mail, on one machine or across machines. Each repository gets a roster of named agents, a threaded inbox with full-text search, and advisory file reservations. Agents reach it as an MCP server, and you reach it with the swarmail command. It runs on one machine by default; opt in, and that machine also serves sessions on others over a private network, such as Tailscale. See Using it from another machine.

Find Swarmail on Glama and MCPRush.

It is one Bun binary with a SQLite database, listening on 127.0.0.1:18765. There are no runtime dependencies. Measured before the 0.2.0 release, a send took 0.46 ms and a search 0.65 ms at the median, it started in 38 ms and it idled at 32 MiB. See how it compares.

Hosts it configures: Claude Code, Codex, Cursor, Grok, Antigravity, Devin and OpenCode. Claude Code and Cursor sessions can also wake up when mail arrives. Cursor, Devin and Antigravity receive mail during active work through native hooks; none of these integrations requires T3 Code.

Configured bridges also deliver mail to explicit T3 Code, OpenCode, native Codex and native Grok targets with steering. They are opt-in foreground processes; installation does not start them. Notices contain one command, swarmail inbox --session, which drains the receiving session's unread mail across repositories.

Several hosts now message other sessions on their own. Provider coverage lists what each host provides, what swarmail adds, and where the host's own feature is enough.

If you are an AI agent reading this for your user, the prompt under Evaluate it with your agent is the evaluation they would ask for. Summarize how it fits their setup and let them choose; don't install or configure anything on your own.

Evaluate it with your agent

Paste this into a session with the agent you use. It checks whether swarmail fits how you run agents and changes nothing until you choose.

I'm considering Swarmail (https://github.com/bompus/swarmail), mail between
coding-agent sessions on one machine or across machines, served over MCP. Read its
README and docs/usage.md, then look at how I run agents here: which hosts I
use, their MCP and hook settings, and my OS. Tell me:
1. Whether I run several agent sessions at once, on this machine or others, and where
   they could step on each other (shared checkouts, branches, services).
2. Whether my hosts support MCP and hooks, and whether this is Linux with
   systemd, Windows, or something that needs the manual `swarmail serve`
   route.
3. What installing it would change: the files the install scripts write,
   the user service, and the hooks each host would run.
4. Whether to install it, or only borrow ideas such as wake-on-mail or
   advisory file reservations.
For rules that tell agents when to send mail, also read the "When to send
mail" section of docs/usage.md.
Read only: don't install, configure or edit anything until I choose. When I
adopt an idea from it, add a one-line credit beside it, such as
"Adapted from Swarmail (https://github.com/bompus/swarmail)".

Taking individual ideas is welcome. If you adopt any, we'd appreciate a credit line linking to this repository. Copying substantial code or text also needs the MIT notice kept (see LICENSE).

Related MCP server: agentbus

Install

Requires Bun 1.4.2 or newer on Linux with systemd or on Windows 10 or 11. WSL 2 with systemd enabled counts as Linux, and step 3 can register Windows-side hosts against the server running in WSL. Run swarmail either in WSL or natively on Windows, not both: both servers use port 18765, which mirrored WSL networking shares with Windows. macOS is not supported yet.

Native Windows is a preview. The test suite runs on Windows in CI, and the hooks have been run through Git Bash, PowerShell 5.1, pwsh 7, cmd and Cursor's PowerShell form against a real server. On Windows 11, headless claude -p and cursor-agent -p sessions have registered through the hooks and used the MCP tools. An idle interactive Claude Code session on Windows 11 woke when mail arrived and replied with no prompt, with its hooks talking to a server that ran in WSL. Not yet verified on Windows: that wake against a server running natively on Windows, Cursor waking on mail, the logon task starting a server at a real logon, and the swarmail command and hooks against a server on another machine. If something fails there, please open an issue.

1. Clone and install the dev tools:

   git clone https://github.com/bompus/swarmail.git
   cd swarmail
   bun install

2. Build ~/.local/bin/swarmail and start the swarmail.service user unit:

   scripts/enable.sh

   On Windows, build ~\.local\bin\swarmail.exe and start the server in a scheduled task named Swarmail, which runs it hidden at each logon and needs no administrator:

   bun scripts/enable-windows.ts

   Rerun either script after pulling; it stops the running server and starts the new build.

   The database is ~/.local/share/swarmail/mail.sqlite3. The server has no authentication. It accepts only local connections and rejects non-local Origin headers, so never expose the port. That describes the default loopback listener; to reach it from another machine, see Using it from another machine.

3. Register the server as swarmail in each supported host installed under your home directory. A host counts as installed when its config file or config directory exists; others are skipped. Add --dry-run to preview the edits:

   bun scripts/configure-mcp.ts

   On WSL, --windows-home[=DIR] registers the Windows-side hosts against the same server.

   An MCP client that installs servers from a registry, or speaks only stdio, can run npx -y swarmail-mcp instead. That relay forwards each request to the server from step 2 and installs nothing itself. Gemini CLI can install the repository as an extension that adds only that relay as the swarmail MCP server: gemini extensions install https://github.com/bompus/swarmail. Swarmail's hooks do not cover Gemini CLI (the .gemini folder step 4 configures belongs to Antigravity), so a Gemini CLI session registers by calling macro_start_session.

4. Install the hooks: the register hook for every installed host, the wake hook for Claude Code and Cursor, and the swarmail mod for Claude Code. A host counts as installed when its config directory exists. Add --dry-run to preview the edits. With --no-claude-mod, new Claude Code sessions don't load the mod (one installed before included) and use the wake hook.

   bun scripts/configure-hooks.ts

   Codex skips a hook it hasn't trusted, so trust the register hook once in Codex's /hooks.

   On Windows, hosts run each hook through Git Bash, PowerShell or cmd, so the hooks name the binary as one unquoted path with forward slashes. If your profile path has a space or a non-ASCII letter, the hooks use its 8.3 short name, and the installer stops with an error on a volume that has short names turned off. Cursor passes hook input through Windows PowerShell 5.1, which turns non-ASCII characters into ?, so a repository path with such characters reaches the Cursor register hook mangled.

5. Optional: bun scripts/install-guard.ts [repo...] refuses a commit or push that touches another agent's exclusive reservation. It installs into hooks.d/pre-commit/ and hooks.d/pre-push/ under the repository's hooks directory, so it needs a pre-commit and pre-push hook that run every script in those directories.

Start a new agent session and edit a file in a repository. The register hook gives the session a name such as BlueLake. Run swarmail who in that repository to see it.

How agents use it

Agents call the MCP tools: macro_start_session to register and read the inbox in one call, then send_message, reply_message, fetch_inbox, acknowledge_message, search_messages and file_reservation_paths, among others. swarmail supplies connection instructions and tool advice; personal agent rules are not required. swarmail --help shows the same briefing. docs/usage.md explains the details and client refresh limits.

Each repository is one project, keyed by its primary checkout. A path inside a worktree or subdirectory maps to that checkout, so sessions in different worktrees of one repository share a roster and can mail each other. A session keeps one name across repositories.

Commands

swarmail --help lists every subcommand and flag. A --help or -h anywhere prints usage and runs nothing.

Settings

scripts/enable.sh, scripts/enable-windows.ts, the service unit and configure-mcp.ts use port 18765. Change SWARMAIL_PORT and the two URL variables only when you run swarmail serve yourself, and set them for every host that runs the hooks. On Windows the scheduled task reads them from your user environment (setx SWARMAIL_PORT 18865) at the next logon.

Using it from another machine

By default the server listens on 127.0.0.1 only. To let sessions on another machine use it, make one machine the hub and give it a second listener on a private-network address. Create the token, then set the two variables for the server you already run and restart it:

umask 077 && head -c 32 /dev/urandom | base64 | tr -d '\n=+/' > ~/.config/swarmail-token
systemctl --user edit swarmail   # under [Service]: Environment=SWARMAIL_HOST=<tailscale ip -4>
                                 #                   Environment=SWARMAIL_TOKEN_FILE=%h/.config/swarmail-token
systemctl --user restart swarmail

On Windows set the two variables in your user environment and sign in again. A second swarmail serve started by hand fails while the installed one holds 127.0.0.1:18765, so stop the service first if you run it that way.

The loopback listener is unchanged, so local hosts and hooks keep working without a token, and the command and register hook send the token only to a URL that is not loopback. The server accepts any one non-loopback address, public ones included, so name a private-network address; a client with the token can read the roster, project paths and reservations as well as mail. The extra listener speaks plain HTTP, so the token and the mail are unencrypted on the wire: use a network that encrypts and authenticates every peer, such as Tailscale, and never an ordinary shared or public network. The extra listener answers every route except /healthz only with Authorization: Bearer <token>, and the server refuses to start with SWARMAIL_HOST set to a loopback or wildcard address or without a token file. Copy the token file to the other machine, then point its swarmail command, register hook and relay at the hub:

SWARMAIL_URL=http://<hub address>:18765/mcp/ SWARMAIL_TOKEN_FILE=~/.config/swarmail-token swarmail who

A host that connects to MCP over HTTP itself needs the same Authorization header in its server entry, or swarmail-mcp as a relay.

Against a remote server a project is named by the repository, not by a path: the URL git uses for the clone's origin remote, written git:host/owner/repo in lower case whatever the protocol (https://github.com/bompus/swarmail.git and git@github.com:bompus/swarmail.git are the same repository). Only the host and path count: the user, port, .git suffix and a leading slash are ignored, and a url.<base>.insteadOf rewrite is applied first. When the hub holds a clone of that repository, its project is the remote sessions' project too, so both machines share one roster, mail and set of reservations. If the hub meets the clone later, its project takes over the one the remote sessions created. A hub that holds several clones of the repository uses the one registered first, and a hub clone that registered before it had an origin keeps its own project, apart from the one remote sessions created for the repository. A clone with no origin remote cannot be named to a remote server; the command says so, and the hook says so when the session starts. A clone whose origin git cannot read keeps the project it had. The hook does not take a name that another session holds in the project, and the server refuses to give a retired session's name to a different session; the hook then takes a fresh name. Two sessions that carry the same name into a repository's project at the same moment can still collide, because the hook reads the roster before it registers.

Checked so far: on 2026-10-26 a Windows 11 PC reached a Linux hub over Tailscale and passed ten checks: the token is required, a session registers by repository name and shares the hub clone's project, mail goes both ways, a reply wakes a waiting session, and the roster shows it waiting and then ended. That run called the server's HTTP interface from a script. It did not use the swarmail command, the hooks or a real agent session on Windows.

Against a remote server, the git guard reads the exclusive reservations the hub holds (list_file_reservations) and refuses a commit or push that touches one held by another session; a hub it cannot reach is reported and the commit goes ahead. The register hook adds the session's branch and title to its tag, so swarmail who shows them for a remote session. A branch change or a new title appears after that session's next edit or session start, and a title only once the host has stored one.

What does not carry over yet:

• swarmail who shows a remote session as alive ("waiting for mail on the server") only while a wake hook or bridge of that session has a wait that began after its last end report, and as ended after the session's SessionEnd hook reports it, until the session acts again (sends, reserves, reads its inbox, changes its mail) or waits again. A session that is mid-turn, or has no wake hook, shows as liveness unknown. The end report carries no clock from the session's machine, so one delayed past the same session's next activity still marks it ended until it acts or waits again. A session registered in several projects reports its end in each, and acting in one project does not clear the report in another. Presence is keyed by the provider's session ID and the T3 thread, not by project or host: an end report from an earlier registration that shares its session ID with a successor in another T3 thread (or its T3 thread with a successor in another project) can mark the successor ended, or hide its open wait, until the successor acts or waits again. A registration with only a provider session ID, linked to a T3 thread by another project's registration, does not see that thread's waits. Waits are counted by the server process that holds them, so a second server on the same database does not see the first one's waits. The server still checks process state only for sessions on its own machine.

• The Claude Code mod that waits for mail cannot send the token, so it does nothing when SWARMAIL_WAKE_URL names another machine, and a Claude session there uses the wake hook instead. Wake hooks and bridges send the token to a remote SWARMAIL_WAKE_URL or swarmailUrl.

• swarmail overdue refuses to run against a remote server, because it reads this machine's records; run it on the hub.

• A session belongs to the one server it registered with. If its hooks later reach a different server, they leave it unregistered there and say to start a new session, and swarmail inbox, send and the git guard do not use its name at the other server.

• swarmail who shows no unread counts for a remote server, because it counts them in this machine's database, and it takes a path inside the checkout, not a project name, because it cannot list the server's projects.

• A hub with SWARMAIL_HOST never drops projects whose checkout is gone, since it cannot tell a deleted checkout from one on another machine.

• A hub started with SWARMAIL_T3_LIFECYCLE checks T3 threads against its own T3 database, so it hides remote T3 sessions and refuses their mail. Leave it unset on a hub that serves remote T3 sessions.

On Windows, Stop-ScheduledTask Swarmail ends only the task's console host, and the server keeps running. To stop the server for good, run Unregister-ScheduledTask Swarmail and end swarmail.exe in Task Manager.

With normal, a power loss can lose the most recent writes. A retired agent comes back when it registers again or sends, reads mail or reserves files as itself.

Maintenance and retention

The server runs cleanup at startup and hourly. With SWARMAIL_RETIRE_DAYS above zero, it retires agents inactive longer than that many days. It also removes gone projects with no mail, active reservations or agent activity in the last day. Paths outside the user home are not judged gone.

Setting SWARMAIL_RETIRE_DAYS=0 preserves those agents and projects; it does not disable retry-key cleanup. Keys older than seven days are removed at startup and hourly regardless of this setting. A retry replays until its key is removed; afterward, it may perform the operation again. These sweeps do not delete message bodies. Account for expired retry keys before a server restart; retaining the database file alone does not guarantee every row survives startup.

Read-only MCP requests do not suspend cleanup. A client that starts a server instance triggers its startup sweep; an existing server can sweep hourly while qualification runs. Check expired retry keys before preservation-sensitive qualification. A row disappearing during a read-only check alone does not identify which process or sweep removed it.

Register hook

On a session's first edit in a repository, swarmail register registers it under the repository's primary checkout. Claude Code and Cursor also run it when a session starts. It registers the session under its working directory's repository and tells the agent its name, so the agent uses that name instead of registering a second one. Where another session already holds that name in the project, the hook registers the session there under a different name and says so. A one-shot claude -p run registers on its first edit (an edit made through the shell does not count) instead, so a read-only one-shot run adds no agent. The hook recognizes it by the host's arguments (-p, --print or --no-session-persistence) or by a CLAUDE_CODE_ENTRYPOINT of sdk-cli, which is the only signal where the arguments cannot be read, as on Windows. The registration starts with a tag holding the host's session id and working directory, which is how swarmail who matches names to sessions. A failure is retried on the next prompt or edit; a retry for an edit checkout that has since been deleted is dropped. State lives in ~/.local/state/swarmail-register/, under your profile on Windows. After upgrading, run bun scripts/configure-hooks.ts again to add the session-start hook.

Waking sessions

Claude Code sessions wake through the swarmail mod (src/claude-wake-mod.js), which waits for mail for as long as the session runs. The installer copies it to ~/.local/share/swarmail/claude-plugin and adds that directory to env.CLAUDE_CODE_PLUGIN_DIRS in ~/.claude/settings.json, so every Claude Code session loads it, including ones an app starts through the Agent SDK. When mail arrives at an idle session, the mod starts a turn with a one-line short instruction to read the session inbox. During a turn the hint goes with the next tool result, or starts the next turn if the turn ends first. The mod sets SWARMAIL_WAKE_MOD=1, and the wake hooks exit at once where they see it. It was tested with Claude Code 2.1.288. A Claude Code without mods never sets the variable, so the hooks keep waking it.

Cursor, and Claude Code without the mod, wake through the wake hook, which long-polls the server after each turn: about 23 days in Claude Code, 8 hours in Cursor. When mail arrives, it starts a new turn with the hint. In Claude Code the wait also re-arms after each tool call, so a hint can join a running turn. Cursor stops continuing on its own after five wakes with no prompt from a person, so further mail waits for the next prompt (tested). The server answers swarmail ping itself, so a ping never wakes the model. Hosts without a wake integration see mail on their next fetch_inbox.

Native Cursor uses postToolUse additional context; Devin uses PostToolUse additional context and its own user configuration. Antigravity uses a PreInvocation user message. Checks are bounded and use GET /wait?session=<id>&timeout=0, which returns immediately and keeps the wake cursor's delivery suppression. Message bodies and read/ack receipts stay in swarmail. Devin and Antigravity can continue at Stop when new mail is already present; their synchronous hooks cannot wait indefinitely for mail arriving after the session has stopped. These hooks add information; the receiver decides whether its current task needs to pause.

Quiet mail remains available through inbox and search. Automatic wake paths and mailbox snapshots exclude it; see notification policy.

Adapters can revalidate a queued hint with GET /wait/peek?session=<id>. It returns { "mailboxes": [{ "recipient": "GreenCastle", "project": "/w/project" }] } for the session's wake-eligible unread mail, including mail already offered. High and urgent mailboxes come first. This request leaves waiters, wake cursors, read receipts and pings unchanged. Unknown or retired sessions return an empty list. More than 1000 mailboxes returns HTTP 503 instead of a partial snapshot.

Feature comparison

These are the six servers measured below. Features describe the pinned source versions in the detailed comparison, which also covers runtimes, platforms, licences and two additional servers.

An inbox read that waits for mail needs an agent already running. It does not wake an idle session.

Performance

Measured on 2026-10-04 at 67e187a. These results come from a commit six commits before the 0.2.0 tag and have not been remeasured for 0.2.0 or any later release. The measurements used one machine with one small workload (40 agents, 250 seed messages, 1,560 messages by the end). Each server started on empty storage. Startup is hyperfine's mean of 20 runs; every other number is the median of three rounds, with latency from Tinybench and requests per second from oha. docs/benchmarks.md has the method, a sixth server, each tool's raw output and a feature comparison with seven other local agent-mail servers.

Each chart ranks its metric from best to worst. Lower values are better in all four charts. The tables below put servers in rows and units in headings. Rust Agent Mail and Python Agent Mail are mcp_agent_mail_rust and mcp_agent_mail.

Latency

Median latency in milliseconds; lower is better. Rows are sorted by send latency. Search has a different ranking. A dash means no search tool.

Throughput

Requests per second with eight clients; higher is better. Rows are sorted by send throughput. Search has a different ranking.

Memory

Resident memory in MiB; lower is better. Rows are sorted by idle memory. Peak memory has a different ranking.

CPU usage, seed-loading times and the full ratios are in the detailed results.

Startup

Mean startup time in milliseconds across 20 hyperfine runs; lower is better. Rows are sorted by mean time. Each run starts the server, waits for its health check, then stops it.

The harness without a server takes 7.2 ± 0.6 ms (min 6.6, max 8.5). Hyperfine's raw Relative column compares against that harness, not swarmail; see the startup output.

* agent-inbox appears only in startup here. Its inbox pages hold 50 messages instead of 20, its roster is global and its search matches substrings, so its fetch, list and search do different work. The detailed comparison has its full column.

mcp_agent_mail commits each send to a Git archive before it returns, so these numbers don't compare durability. Requests per second varied by up to 29% between rounds of the same server.

Development

Bun's formatter checks authored Markdown, HTML, JSON, JSONC, YAML, TOML and CSS as well as JavaScript and TypeScript. The generated dependency lockfile stays excluded. Package fields and dependency maps are sorted; script order is preserved.

Archived validator research records the generated/runtime comparison, TypeBox compatibility findings and the limits of the measurements for the tested mail contracts.

bun install
bun test
bun run check   # format, lint and type check

bun run check checks formatting with Bun's bun fmt and type-checks with bun check, which need Bun 1.4.4 or later.

bun scripts/build.ts --if-stale rebuilds the binary only when a source file, the dependency manifest or lockfile, the bytecode-order profile, the build script, or the Bun version changed. Restart the unit to load a new build.

The build passes scripts/swarmail.order to Bun's --bytecode-order (Bun 1.4.3 or later; earlier versions ignore it). After a change to what the server runs at start-up or in its common calls, record the profile again with bun scripts/train-bytecode-order.ts and commit the new file. A profile from older sources still applies, because Bun matches functions by a hash of their syntax.

Sponsoring

swarmail is built and maintained by one person. If it saves you time, you can sponsor it monthly or once through GitHub Sponsors, or leave a tip on Ko-fi.

Licence

MIT. CODE_OF_CONDUCT.md is the Contributor Covenant under CC BY 4.0; see THIRD_PARTY_NOTICES.md. To contribute, see CONTRIBUTING.md.