openagent.email

Self-hosted email for AI agents. The open-source alternative to AgentMail.

openagent.email · website: openagentemail/website

One docker compose up on your own VPS gives every agent you run unlimited real mailboxes on your own domain — over REST and MCP — with OTP and verification-link extraction built in. No per-inbox pricing, no third party ever seeing your mail.

Quickstart

npx -y @openagentemail/setup

A guided wizard: it checks what you already have, helps you pick a VPS and a domain if you're missing either, and connects your agent clients (Claude Code, Cursor, Kimi Code…) once the server is up.

The manual path needs a VPS with outbound/inbound port 25 open and a domain you control:

git clone https://github.com/openagentemail/openagentemail.git && cd openagentemail
cp .env.example .env   # set DOMAIN, API_KEYS, mailbox password, and NTFY_ADMIN_PASSWORD
docker compose up -d
sudo ./deploy/dns-records.sh   # prints the exact DNS records to create

Then verify everything end to end:

sudo ./deploy/doctor.sh

doctor.sh checks .env permissions; MX, A, SPF, DKIM, and DMARC; PTR; outbound port 25; DNS blocklists; TLS certificates on 465 and 993; and the server-side ntfy verification endpoint. It does not log in over IMAP/SMTP or send a round-trip test. Docker Mailserver can create docker-data/ as root, so use sudo for these two scripts; an EACCES failure prints the same retry instruction instead of pretending the DKIM key is missing.

If no SMTP relay is configured and outbound port 25 is blocked, API queued:true only means the local mailserver accepted the message. It does not mean the recipient received it: Postfix can retain the message in its queue. Treat doctor's outbound-port-25 result as the delivery prerequisite, or configure a relay before relying on direct delivery.

Public TLS with Let's Encrypt (opt-in)

The default docker compose up -d path remains self-signed: it does not start or pull Certbot and does not publish TCP 80. To use a publicly trusted mail certificate, opt in only after mail.$DOMAIN has an A (and, if used, AAAA) record pointing at this host and the firewall permits inbound TCP 80. HTTP-01 cannot create those DNS or firewall prerequisites for you.

In .env, set the following (use a reachable contact address outside this mailserver when possible):

SSL_TYPE=letsencrypt
SSL_DOMAIN=mail.example.com       # exactly mail.$DOMAIN
LETSENCRYPT_EMAIL=admin@example.net  # optional, but recommended

First issue the certificate with the explicitly enabled sidecar; do not start the mailserver in letsencrypt mode before this succeeds:

docker compose --profile letsencrypt-bootstrap up -d certbot-bootstrap
docker compose logs -f certbot-bootstrap
# Wait for “Successfully received certificate”, then confirm:
docker compose --profile letsencrypt-bootstrap run --rm --no-deps \
  --entrypoint ls certbot-bootstrap -l \
  /etc/letsencrypt/live/mail.example.com/fullchain.pem \
  /etc/letsencrypt/live/mail.example.com/privkey.pem

This temporary container reads the shared certificate volume, so confirmation still works after the one-shot bootstrap has stopped.

If first issuance fails, Certbot stops instead of retrying the ACME request in a tight loop. Correct the DNS/port-80/domain prerequisite, then explicitly run the same docker compose --profile letsencrypt-bootstrap up -d certbot-bootstrap command again.

The entire /etc/letsencrypt tree is a persistent named volume shared with the mailserver read-only: Certbot's live/ files are symlinks into archive/, so mounting only live/ is incorrect. Once the first certificate exists, start the full opt-in stack and verify the public endpoints. Do not enable letsencrypt-bootstrap and letsencrypt together: both publish host TCP 80.

docker compose --profile letsencrypt up -d
./deploy/doctor.sh
openssl s_client -connect mail.example.com:465 -servername mail.example.com </dev/null \
  2>/dev/null | openssl x509 -noout -issuer -subject -dates
openssl s_client -connect mail.example.com:993 -servername mail.example.com </dev/null \
  2>/dev/null | openssl x509 -noout -issuer -subject -dates

After bootstrap, the renewal sidecar runs renew every 12 hours and restarts with Docker. docker-mailserver's change-detection service watches SSL_TYPE=letsencrypt certificate updates and reloads Postfix and Dovecot, so renewed certificates take effect on 465/993 without a manual container restart. Keep the letsencrypt profile enabled for normal operation. If it is omitted, the sidecars and TCP 80 are absent and the original self-signed path is unchanged.

Create an identity and hand your agent its scoped token (shown once):

curl -X POST http://localhost:3100/v1/identities \
  -H "Authorization: Bearer $API_KEY" -H "Content-Type: application/json" \
  -d '{"name":"signup-bot","scopes":["read:messages"]}'
# → 201 {"address":"fox-k7d2@example.com","name":"signup-bot","token":"oa_…","scopes":["read:messages"]}

The API binds to 127.0.0.1 by default — reach it from other hosts over an SSH tunnel or a TLS proxy: docs/security.md.

Paths are exact: call /v1/notify, not /v1/notify/ — a trailing slash returns a plain 404 rather than the API error format.

Using your own mail server

Already have a mail provider for your domain? Run the API by itself with compose.api-only.yaml, connected to that provider's catch-all mailbox. The external mail server guide covers the required catch-all setup, Portainer deployment, SMTP sender limits, and TLS certificate verification.

The standalone default project name is openagentemail. If the full compose.yaml stack also runs on the same host, the API-only stack must not share that default project: give it an explicitly different -p value or COMPOSE_PROJECT_NAME so the two stacks cannot adopt each other's resources.

To run multiple API-only instances on one host, give every instance its own environment file, unique Compose project, and host API_PORT. The API always listens on port 3100 inside its container; API_PORT changes only the host-side mapping. For example:

mkdir -p ../oae-api-only-env
cp .env.api-only.example ../oae-api-only-env/alpha.env
cp .env.api-only.example ../oae-api-only-env/beta.env
chmod 600 ../oae-api-only-env/*.env
# Set API_PORT=3100 in alpha.env and API_PORT=3101 in beta.env.
# Generate separate API_KEYS and TASK_SIGNING_SECRET values in each file.

docker compose -p oae-alpha --env-file ../oae-api-only-env/alpha.env -f compose.api-only.yaml up -d
docker compose -p oae-beta  --env-file ../oae-api-only-env/beta.env  -f compose.api-only.yaml up -d

You may set a unique COMPOSE_PROJECT_NAME for each command instead of using -p. The project names make Compose generate distinct container names and project-scoped named volumes (for example oae-alpha_api-data and oae-beta_api-data). Do not reuse a project name or API_PORT between the two instances. Each instance must have independently generated API_KEYS and TASK_SIGNING_SECRET values; configure its IMAP and SMTP credentials for that instance's intended mailbox/provider boundary as well. Keep these populated environment files outside the repository, as in the example above.

Read mail in a browser

Open http://localhost:3100/ui and paste an admin or identity API token. The built-in dashboard lists the addresses the token may access, shows Inbox / All Mail (IMAP) and Sent (API/MCP send audit for 30 days; direct SMTP is not listed) with cursor paging, extracts verification codes and links at the top of a message, offers Rendered (isolated HTML iframe), Plain text, or Source views, and can mark messages read or unread. Source is fetched on demand from a size-capped no-store endpoint and never injected as HTML. Scheduled and Trash are omitted until the backend can serve them. Admin sessions can also create identities (with custom address prefixes), rotate tokens, and delete identities directly from the overview table.

The browser exchanges the token once for an HttpOnly session cookie; manually pasted tokens never enter the URL or browser storage. You can also bookmark https://myinstance:3100/ui?token=<admin-token> for direct login. The server automatically issues a single-use exchange code (TTL ≤ 10 min) via a 302 redirect and sanitizes the URL with Cache-Control: no-store and Referrer-Policy: no-referrer, followed by client history.replaceState cleanup. Tokens passed in URLs can still linger in upstream reverse proxy access logs before redirection — see docs/security.md for reverse-proxy scrubbing guidance. Percent-encode the token if it contains URL-reserved characters like +, &, or # (e.g. a+b → a%2Bb), as + decodes to a space and & truncates the value. Only open ?token= links you generated yourself; if you suspect a link has leaked, rotate the token from the UI immediately. Do not open ?token= links sent by others — the link signs you into the sender's session (the app displays a visible "Signed in via link" banner across the interface for the session as the tell). Sessions live only in API process memory, so restarting the API signs every browser out. They expire after 12 idle hours or 24 hours total — or tick Trust this device at login to keep a sliding 30-day session on that browser. Each token holds at most five sessions; a sixth login evicts that token's least-recently-used one instead of locking you out.

For another computer, use the same SSH tunnel recommended for the API or put a TLS reverse proxy in front. The login form refuses non-local plain HTTP, and session cookies are Secure away from localhost. Set UI_ENABLED=false in .env to make every /ui route return 404.

HTML email is treated as hostile input. The UI removes scripts, images, forms, links, sender CSS, and all attributes except numeric table spans, then loads the result in a separately sandboxed frame with a restrictive CSP. The sanitize-html 2.x dependency is deliberately pinned to an exact version; upgrade it in a dedicated change and rerun the full poison-message corpus.

Admin overview

An admin session lands on Overview: every identity in one table with the message count, unseen count, last delivery, and creation day, plus totals across the top. Each row also shows whether the identity has a token (green dot) and has Rotate and Delete action buttons. A Create Identity button above the table opens a form where you can set a custom address prefix (e.g. qa-bot) or leave it blank for a random one; the new token is shown once in a copy-to-clipboard modal. Identity sessions never see the overview or management controls — they go straight to their own inbox. The page is served from the same in-process API as the rest of /ui; there is no new public endpoint outside /ui/api.

What the numbers mean, and where they stop:

• Counts are a window, not a lifetime total. One scan reads the newest 500 messages in the catch-all mailbox and attributes each one to the identities it was delivered to. The header says newest N of M in the mailbox so the window is never mistaken for history. A message addressed to two identities counts once for each row and once — not twice — in the totals.
• Honest instead of round. Messages with enormous recipient lists can exceed the scanner's per-message and global memory bounds. Rows the scanner could not fully account for show ≥N or Unknown rather than a confident wrong number, the page explains why, and unmatchedInWindow is reported as null instead of a made-up zero. The same applies to an identity created after the last scan: it reads Unknown until the next one, never a false 0.
• Snapshots are cached in memory for 15 seconds and reused for up to 10 minutes while a refresh runs in the background, so opening Overview or walking in and out of inboxes does not hammer IMAP. Refresh is floored at 5 seconds. Restarting the API drops the cache — the first request afterwards pays for a fresh scan.
• Failures cool down and never lie. If a scan fails, the next attempt waits 5 seconds (the API sends Retry-After), the table keeps showing the previous numbers, and the header says the last refresh failed. Once a snapshot is older than 10 minutes it is not revived by a failed refresh: the page reports the counts as unavailable instead of showing stale data as current. While a cold scan is still running the endpoint answers 202 and the address list renders immediately with Loading… in the count columns.
• GET /ui/api/overview (browser session only, admin only) returns exactly the fields the page renders — never message content. ?refresh=1 asks for a new scan, subject to the 5-second floor and the failure cooldown.

Deliberate limits, so nothing here is a surprise later:

• Overview shows counts and timestamps only. Subjects, senders, and verification codes need per-message parsing, which is what the inbox view is for.
• New mail can be up to 15 seconds late on the page; Refresh fetches sooner. There is no steady-state polling: the page only schedules a follow-up while counts are loading or a refresh is pending, capped at 15 attempts over 20 seconds and paced by the server's own retry hint.
• A scan that misses its deadline is abandoned even if IMAP answers a moment later, and the next request scans again. The deadline covers connecting as well as fetching, so a hung server does not park a request behind IMAP's own 30-second socket timeout.
• Up to 200 identities render in one pass. Beyond that, expect to want paging or virtual scrolling; filtering and sorting happen in the browser today.
• The dashboard self-hosts the Satoshi webfont (/ui/fonts/*, the same typeface as the website) so it renders identically on every machine; font-src is 'self'. The favicon is an SVG (/ui/favicon.svg) so it needs no build step, and /ui/favicon.ico keeps returning 204 as before.
• Form controls use a dedicated --line-control border token so their outlines stay above 3:1 contrast. It is the one intentional deviation from the website's palette and is a one-line revert.

Features

• Unlimited identities — one catch-all mailbox, unlimited anything@yourdomain addresses. No provisioning, no per-inbox cost.
• Scoped tokens — every identity gets its own token that can only read and send as that address. The admin key never has to touch your agents.
• REST + MCP — the same operations over a plain HTTP API and a first-class MCP server your agents can call directly.
• Server-side notifications — private ntfy transport for human alerts and managed-agent wake-ups, with OTP-only mail notifications by default. Topics and ntfy credentials stay on the server; phone setup is deliberately a later v0.3.1 step. Per-identity push content tiers control how much of each mail-arrival alert leaves the server: tier 1 interrupt only (default), tier 2 adds subject/from, tier 3 (admin + explicit risk confirm) adds a short body preview and extracted OTP codes/links.
• mail_wait_for / POST /v1/messages/wait — long-poll an inbox until a matching message arrives, with OTP codes and verification links already extracted. Built for automated signups.
• Read/unread state — mail_mark_seen / POST /v1/messages/:id/seen lets an agent (or the human in the dashboard) mark a message handled, so the unseen count means "still needs attention". An identity may flag mail it received or that this server actually sent (TO ∨ trusted Sent: From match and Message-ID in the outbound registry), matching the Sent folder (#26 PR 2); a forged From does not count. It cannot flag another identity's mail. Reading a message never changes the flag by itself.
• Web dashboard for humans — inspect identities and messages at /ui (Inbox is the default landing for every session). Inbox is a three-pane mail client (identity/folder, list, detail) with Rendered / Plain text / Source; HTML stays in a sandboxed iframe. Real /ui/* History routes cover Overview, Tasks, Notifications, and Configure; the shell stays a zero-bundler /ui/styles.css + /ui/app.js pair.
• Safety rails built in — per-identity send rate limits (20/hour default), automatic mail retention (30 days default), localhost-only API binding.
• Bring your own relay — send directly from the VPS, or route outbound through Amazon SES / SMTP2GO / any SMTP relay with one env var.
• DNS wizard + doctor — deploy/dns-records.sh generates your exact DNS records; deploy/doctor.sh diagnoses deliverability before your agents depend on it.
• Single dependency: Docker. The stack is the API, docker-mailserver, and a private ntfy container. Nothing else.

How it works

┌─────────────┐   MCP (stdio)    ┌──────────────────┐
│  AI agents   ├──────────────────▶                  │
│ (Claude Code,│                  │   openagent api  │
│  Cursor, …)  │   REST /v1/*     │  (Node, imapflow │
└─────────────┘──────────────────▶ │   + nodemailer)  │
                                   └────────┬─────────┘
                                            │ IMAP + SMTP (localhost)
                                   ┌────────▼─────────┐      SMTP 25
                                   │ docker-mailserver│ ◀──────────▶  the world
                                   │ catch-all mailbox│  (or your relay: SES, …)
                                   └──────────────────┘

One catch-all account on your domain receives everything. The API logs into it over IMAP, matches messages to identities by the To/Delivered-To header, and sends via SMTP with the From rewritten to the chosen identity. Polling + IMAP IDLE for low-latency waits.

Multi-domain support

You can configure secondary domains alongside the primary DOMAIN by setting EXTRA_DOMAINS=domain2.com,domain3.org in .env.

• Identities: Agents can be created on the primary domain or any configured secondary domain via the API (POST /v1/identities with domain), MCP tools (mail_new_identity), or the Web Dashboard modal. Note that creating the same localpart across different domains is not allowed for now and will be revisited in Part 2.
• Inbound & Outbound: Catch-all inbound matching and internal mail stamping (X-OA-Mail-Stamp) recognize all configured domains.
• Tasks & Alerts: Cross-domain tasks and notifications seamlessly support agents across all configured domains, with identical localparts across domains currently not allowed (409) pending a design revision, and agent-card endpoints still publishing the primary domain.
• Pending Deploy Automation: Automated deploy scripts (deploy/dns-records.sh, deploy/doctor.sh, Certbot SANs, and multi-domain DKIM key generation) currently configure the primary DOMAIN. Automated multi-domain provisioning scripts will follow in Part 2.

Optional compliance archive

Set ALWAYS_BCC=archive@example.net only when your compliance policy permits an additional external delivery copy. It is off by default and adds the archive once to the SMTP envelope for API, MCP, and task sends, preserving visible recipient order and matching an existing recipient only by exact local-part and case-insensitive domain—without adding a MIME Bcc header or changing visible To, header From, envelope MAIL FROM, SPF, DKIM content, or DMARC alignment. Archive mailbox access, privacy, retention, aliases, and forwarding are operator/MTA responsibilities.

The archive is an independent, off-domain compliance destination—not an ordinary untrusted recipient. For all-local visible recipients it receives the exact MIME, including the X-OA-Mail-Stamp that preserves local internal classification. Same-domain ALWAYS_BCC values are rejected at startup: on a shared catch-all deployment they are not an independent archive destination, and duplicate suppression is not a substitute for that boundary. Configure a controlled external compliance mailbox or leave ALWAYS_BCC unset.

Every enabled archive requires an explicit TASK_SIGNING_SECRET of at least 32 characters. The application enforces only presence and that length minimum; it does not prove entropy, so operators must generate a high-entropy secret (the Compose examples use openssl rand -hex 32). The historical SMTP-password fallback remains only when the archive is absent or blank. Both Compose deployments already require the dedicated secret.

If at least one original recipient is accepted and the configured archive RCPT is rejected, the API preserves Nodemailer's partial-success semantics and logs a content-free warning, even when another original recipient was rejected. If the archive is accepted while no original recipient is reported accepted, the API fails instead. After an upstream SMTP server accepts or queues the archive RCPT, later delivery failures appear in SMTP/Postfix/relay logs or DSNs rather than synchronously in the API response.

SMTP result matching deliberately keeps local-part spelling exact. A relay that rewrites accepted RCPT local-part case can therefore cause a conservative failure/retry; configure relays to preserve RCPT spelling. Case-distinct archive and original local-parts are intentionally separate SMTP mailboxes and can produce two copies with a case-folding provider, so use consistent spelling.

Use it from your agent (MCP)

Requires Node.js 18+ on the machine running the MCP client — no install step, npx downloads and runs the package on first use. The stdio client only needs OPENAGENTEMAIL_API_URL and OPENAGENTEMAIL_API_KEY; it does not read the mail server's DOMAIN / API_KEYS / IMAP / SMTP variables.

claude mcp add openagentemail \
  --env OPENAGENTEMAIL_API_URL=http://localhost:3100 \
  --env OPENAGENTEMAIL_API_KEY=oa_your-identity-token \
  -- npx -y @openagentemail/mcp

Or the raw JSON config (Claude Desktop, Cursor, Kimi Code):

{
  "mcpServers": {
    "openagentemail": {
      "command": "npx",
      "args": ["-y", "@openagentemail/mcp"],
      "env": {
        "OPENAGENTEMAIL_API_URL": "http://localhost:3100",
        "OPENAGENTEMAIL_API_KEY": "oa_your-identity-token"
      }
    }
  }
}

Tools

Typed approval actions are JSON-only and limited to 65,536 canonical UTF-8 bytes, root-inclusive depth 10, and a server-clock lifetime of 30 days. Exact limits pass; REST and MCP surface approval_action_too_large, approval_action_too_deep, or approval_expiry_too_far as stable client errors when a bound is exceeded.

The exact v1 action-digest recipe and public cross-runtime vectors are in docs/approval-digest.md.

Task leases are opt-in; TASK_LEASES_ENABLED defaults to false, so existing clients remain compatible. If the flag is turned off after a lease exists, list/detail responses retain its safe timing and generation fields and add leaseStatus: "disabled"; no lease is silently cleaned up. A lease generation is capped at 24 hours from its initial claim, and no claim or renewal is allowed at or after seven days from the task's first claim; renewals cap their deadline rather than resetting either anchor, and equality is rejected. During an active recipient lease, omitting the optional credential retains task_already_terminal; a supplied wrong, malformed, expired, or reclaimed-generation credential returns task_lease_required at HTTP 409. The opaque leaseToken is only the claim bearer and is never listed, rendered, logged, or emailed. When TASK_LEASES_ENABLED=true, each mailbox requires exactly one API process (single-replica deploy). Multi-process sharing of a mailbox is an unsupported boundary: concurrent claims can assign the same generation, and durable rebuild fail-closes on same-generation different-body claims (the task is hidden; this is not silent corruption and has no cross-task impact). Future cross-process coordination, if introduced, will be derived from the durable stream under the already-decided redesign in issues #80 and #84. Optional TASK_LEASES_EXPIRY_AUDIT_M3 (default false) decouples reclaim from expiry-audit SMTP: when on, claim derives lease expiry from server time so audit delivery cannot block a new claim. Late matching expiry receipts are always rebuild no-ops (ungated). Optional TASK_LEASES_OVERLAY_BOUND (default false) stops public list/detail replay of unindexed lease overlay events after 15 minutes from sentAt.

Full per-client setup (Claude Code, Claude Desktop, Cursor, Kimi Code, generic): docs/mcp-clients.md · server details: packages/mcp/README.md

Why self-host?

• Privacy — OTP codes and verification links are credentials. Self-hosted, they never leave a machine you own. No third party reads, stores, or trains on your mail.
• Cost — a $5 VPS and a domain you already have vs. per-inbox/per-message SaaS pricing that scales linearly with your agent fleet.
• Control — your IPs, your reputation, your retention. No rate limits, no account suspensions, no sudden API deprecations.

Server requirements

Measured on our own production instance, idle: ~190 MB RAM total, ~0% CPU, and ~2 GB of disk for the Docker images. Mail itself is a rounding error — retention auto-deletes after 30 days.

That's a $5/mo VPS — or a $10–15/year deal box. The real prerequisite isn't size, it's port 25: AWS, GCP, Azure, DigitalOcean and Vultr block it by default (some unblock on request). Check before you buy — or route outbound through a relay and you don't need port 25 out at all.

Comparison

* Push tiers 2/3 may relay subject/from or body/OTP via ntfy (off by default). † AgentMail Outposts: email content stays in your AWS account; AgentMail still runs dashboard, auth, billing, and upgrades. BYOC ≠ open-source self-host on any VPS.

Roadmap

• v0.1 — REST + MCP, catch-all identities, wait_for with OTP/link extraction, DNS wizard + doctor, optional SMTP relay.
• v0.2 — scoped per-identity tokens, send rate limits, automatic retention, localhost-safe defaults, expanded OTP corpus.
• v0.3 (current) — built-in private ntfy notifications, OTP-aware IMAP watcher, server-side agent wake-ups and notification ACLs. Phone delivery and webhooks are intentionally out of this first release.
• Distribution — planned one-click app in the OpenShip catalog.

Docs

• docs/api.md — REST API reference with curl examples
• docs/security.md — tokens, exposure, rate limits, retention
• docs/mcp-clients.md — MCP setup for every major client
• docs/cloudflare-zero-trust.md — Cloudflare Tunnel and Access for the Dashboard without blocking MCP
• docs/agent-signup.md — let agents finish sign-ups that email a code, with a verified OKX wallet example
• docs/dns-setup.md — DNS records, explained one by one
• docs/deliverability.md — the field guide to actually landing in the inbox

Contributing

Issues and PRs welcome — see CONTRIBUTING.md.

License

Apache-2.0