MCP Ecc

mcp-ecc

Email · Calendar · Contacts — one MCP server for all your accounts

A Model Context Protocol (MCP) server that lets AI assistants read, write and manage your email, calendar and contacts — across Google, Microsoft 365/Outlook, Zoho, and any IMAP/SMTP, CalDAV or CardDAV account — from a single interface.

---

mcp-ecc aggregates your mail, calendar and contacts into standardised MCP tools (mail.*, calendar.*, contacts.*, accounts.*) so any MCP client — Claude, Cursor, and others — can work with all of them uniformly.

• One OAuth consent per cloud provider covers mail + calendar + contacts
• Multi-user — users self-manage their own accounts, each with a per-user MCP API key
• Credentials encrypted at rest (AES-256-GCM)
• Self-hostable — single container serving a web admin UI, REST API and the MCP endpoint on one port

---

Quick start

Docker

docker run -d --name mcp-ecc \

-p 127.0.0.1:3001:3001 \ -v mcp-ecc-data:/data \ karljsamuel/mcp-ecc:latest

Publish the port on 127.0.0.1 as above. The container listens on 0.0.0.0 internally so the published port can reach it — the loopback restriction on the *host* side is what keeps it off your network. Put a reverse proxy or tunnel in front before widening that.

With no configuration the container starts in setup mode. Read the one-time setup token from the logs:

docker logs mcp-ecc 2>&1 | grep "Setup token:"

Then open http://127.0.0.1:3001/setup, paste the token, and work through the wizard: choose SQLite or Cloudflare D1, let it verify the connection, generate an encryption key, set the server options, and create the first admin account.

While setup is incomplete, every API route except /api/setup/* answers 503. The setup token is the only thing protecting that page until an admin exists, so do not publish the port beyond loopback until setup has completed.

Volume permissions (bind mounts)

The container runs as uid 1001 (the mcp-ecc user), not root. A named volume is seeded with that ownership automatically. A bind mount is not — Docker mounts the host directory exactly as it is, so if it is owned by root the app can read it but cannot write, and you get EACCES: permission denied on /data (or a database that never initialises).

If you use a bind mount — as docker-compose.yml does with ./data:/data — prepare the directory with the container's uid before starting:

mkdir -p ./data

sudo chown -R 1001:1001 ./data sudo chmod 2775 ./data # optional: keeps group access for host users

Then confirm from inside the container:

docker exec mcp-ecc sh -c 'touch /data/.probe && rm /data/.probe && echo writable'

Symptoms of getting this wrong: a permission denied line in docker logs, a crash/restart loop, and a 502 from your reverse proxy or tunnel because nothing ends up listening on the published port.

The .env must sit beside docker-compose.yml

Compose interpolates ${MCP_ENCRYPTION_KEY}, ${CLOUDFLARE_API_TOKEN} and friends from a .env in the same directory as the compose file. If that file is missing or misnamed, every variable resolves to an empty string and the container starts in setup mode — which looks like an outage: /health returns 200 while /mcp and /api/* return 503, and the logs say is unconfigured and running in SETUP MODE.

Check that the file is where compose expects it, and that the keys are actually present (this prints variable names only, never values):

ls -la /path/to/compose-dir/.env

grep -oE '^[A-Z_]+=' /path/to/compose-dir/.env

Note that environment variables are captured when a container is created. If you edit .env or rotate a credential, recreate the container — docker compose up -d --force-recreate — or the running container keeps the old values.

The CLI is on PATH inside the container

The image links mcp-ecc into /usr/local/bin (→ /app/packages/cli/dist/bin.js), so the CLI is available by name with nothing installed on the host — for stdio MCP clients, maintenance, and the setup flow:

docker exec -it mcp-ecc mcp-ecc --help

Persisting the generated configuration

A .env written inside a container is lost when the container is recreated, so the wizard hands you the content through a copy button instead of writing it. Put it somewhere durable and give it to the container by reference:

# paste the copied content into ./mcp-ecc.env first

docker rm -f mcp-ecc docker run -d --name mcp-ecc \ -p 127.0.0.1:3001:3001 \ -v mcp-ecc-data:/data \ --env-file ./mcp-ecc.env \ karljsamuel/mcp-ecc:latest

Alternatively, mount a writable location and point MCP_ENV_FILE at it. The wizard then writes the file itself — no copy step, no restart — because the path is known to survive:

docker run -d --name mcp-ecc \

-p 127.0.0.1:3001:3001 \ -v mcp-ecc-data:/data \ -e MCP_ENV_FILE=/data/.env \ karljsamuel/mcp-ecc:latest

Behind Cloudflare Tunnel or another reverse proxy

Set PUBLIC_URL to the public address — OAuth callbacks and the MCP endpoint URL are built from it, so localhost will break them:

-e PUBLIC_URL=https://mcp-ecc.example.com

If you put Cloudflare Access in front, leave a bypass for /mcp: MCP clients authenticate with per-user Bearer keys and cannot complete an interactive login.

Web UI

mcp-ecc setup          # or: start the API unconfigured and open /setup

Either path walks the same steps: pick SQLite or Cloudflare D1, verify the connection actually works, generate an encryption key, set the server options, write .env, then create the first admin account.

CLI (npm)

npm install -g mcp-ecc

mcp-ecc setup # first run — writes .env and creates the admin account mcp-ecc # interactive TUI — login, add accounts, manage everything mcp-ecc start # stdio MCP server for agent hosts

From source

git clone https://github.com/karljsamuel/mcp-ecc.git

cd mcp-ecc npm install npm run build node packages/cli/dist/bin.js --help

---

Upgrading to 0.7.0

This release invalidates stored credentials. Every account must be re-authenticated once after upgrading, and per-user API keys are regenerated. Back up your database first if you want a rollback path.

What changed

Credential encryption was rebuilt from scratch:

• One canonical ciphertext format shared by both backends: base64(iv[12] || authTag[16] || ciphertext), AES-256-GCM.
• Per-owner key derivation. The data key is derived with HKDF-SHA256 over MCP_ENCRYPTION_KEY using the owner id, instead of one global key. A value encrypted for one user cannot be decrypted as another.
• Key rotation is supported. Each encrypted row records a key generation in the database, so you can move to a new key and re-encrypt in batches while both generations remain readable.
• CryptoJS is gone — the dependency, the legacy reader, and the migration path with it. There is no automatic migration from pre-0.7 ciphertext.
• No fallback key. A missing or placeholder MCP_ENCRYPTION_KEY is a hard startup error. The key is never read from or written to the database.

Why you must re-authenticate

Values written by earlier versions cannot be decrypted by this one, by design. Everything encrypted at rest — OAuth refresh tokens, OAuth client secrets, IMAP/SMTP app passwords — becomes unreadable. You can start from a fresh database, or upgrade in place and re-run the auth flow per account.

If you were told to run 0.6.x first

That instruction is obsolete. 0.6.x was an intermediate step toward a platform-crypto format; 0.7.0 supersedes it and removes both the intermediate step and the reader it depended on. Going straight from 0.6.0 to 0.7.0 is the supported path.

Rotating the key

1. Set MCP_ENCRYPTION_KEY to the new key.
2. List the old key as MCP_ENCRYPTION_KEY_PREVIOUS in the form " : " (comma or whitespace separated for several).
3. Call rotateEncryptionKeys(batchSize) on the storage adapter — rows are re-encrypted in batches and both generations decrypt during the changeover.
4. Once every row is rotated, drop MCP_ENCRYPTION_KEY_PREVIOUS.

Set MCP_ENCRYPTION_KEY_GENERATION explicitly when you rotate, so a later key is not renumbered into an earlier generation's slot.

Connecting an MCP client

Point your client at the HTTP endpoint with your per-user API key (shown in the web UI under Settings):

{

"mcpServers": { "mcp-ecc": { "type": "http", "url": "http://localhost:3001/mcp", "headers": { "Authorization": "Bearer " } } } }

Agents can read SKILL.md or llms.txt for guided, automated setup.

---

Providers

Microsoft 365: app passwords are being retired — use OAuth (Microsoft Graph) only.

CalDAV/CardDAV compatibility: the providers implement standard WebDAV (CalDAV RFC 4791, CardDAV RFC 6352) with Basic auth and should work with any standards-compliant server. Tested against Radicale 3.7.8. Not tested against other servers; OAuth-only providers (e.g. iCloud without an app-specific password) are not supported.

Deployment notes: Google/Microsoft/Zoho/IMAP/SMTP/CalDAV/CardDAV run under Node.js (CLI or Docker). Cloudflare D1 can be used as a hosted alternative for the local SQLite database.

---

MCP tools

mail.listFolders · mail.listMessages · mail.getMessage · mail.sendMessage

mail.searchMessages · mail.moveMessage · mail.setFlags · mail.deleteMessage

calendar.listCalendars · calendar.listEvents · calendar.getEvent · calendar.freeBusy calendar.createEvent · calendar.updateEvent · calendar.deleteEvent

contacts.list · contacts.get · contacts.create · contacts.update contacts.delete · contacts.search

accounts.list · accounts.get · accounts.add · accounts.remove · accounts.sync

Full reference with input schemas: docs/mcp-tools.md

---

Documentation

• [Getting started](https://github.com/karljsamuel/mcp-ecc/blob/main/docs/README.md) — overview and quick start
• [CLI reference](https://github.com/karljsamuel/mcp-ecc/blob/main/docs/cli.md) — every mcp-ecc command and workflow
• [Auth & users](https://github.com/karljsamuel/mcp-ecc/blob/main/docs/auth-users.md) — multi-user model,