Agent Relay
io.github.mmmikaelPublished by GitHub account mmmikaelv0.2.0Updated Oct 24, 2026
End-to-end encrypted inbox where AI agents deliver files and request human approval.
Overview
Agent Relay lets an AI agent deliver encrypted files and email drafts into a private inbox where a human reviews and approves them before anything is sent.
- What you can do
- Deliver work into an end-to-end encrypted inbox, with artifact preview and download for text, Markdown, HTML, PDF and images (deliver_to_inbox). List inbox sessions so an agent can check what it has already delivered (list_inbox_sessions). Submit encrypted outbound email drafts that a human previews, then approves or rejects before sending (submit_email_draft). Verify the token and encryption setup from the command line before wiring it into an agent.
- When to use it
- Use it when an agent produces finished work, drafts emails, or requests spend and you want a human approval gate before anything leaves, sends, or charges. It suits Claude Code, Cursor, Claude Desktop or any HTTP-capable client that needs a private delivery channel.
- Requirements
- Runs locally over stdio via the npm package @arelay/cli (Node.js). Needs an Agent Relay account with passkey sign-in and completed encryption setup, plus an agent token set in ARELAY_TOKEN; ARELAY_URL points at a self-hosted deployment and defaults to the hosted service. Network access to the relay is required.
Installation
In SourceWeft
- Open Agent Relay in the dashboard and add it to a workspace.
- Enable the server for the chats that should use its tools.
Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.
Other MCP clients
Follow the launch instructions in the repository.
README
Agent Relay
[CI] [CodeQL] [License: MIT] [Node.js] [TypeScript] [Svelte] [Website]
Agent Relay is where your AI agents ask permission. Agents draft emails, request spend, and deliver finished work into a private, end-to-end encrypted inbox; nothing leaves, sends, or charges until you approve it. Anything that can make HTTP requests can use it — Cursor, Codex, Claude Code, cron jobs, webhooks, or your own backend — and it is open source under MIT, so you can run it yourself.
[An encrypted delivery arriving in the Agent Relay inbox and the report opening in the preview pane]
An optional Email Review Relay plugin adds a human-in-the-loop outbound email path: agents submit
encrypted drafts, you preview them in the same inbox, then approve or reject before anything
is sent. It is enabled on arelay.app; self-hosters can turn it on with
EMAIL_REVIEW_RELAY_ENABLED=true (setup).
Two ways to use it:
- arelay.app — hosted service; sign up and connect your agents (no deployment).
- Self-host — run this repo on your own infrastructure under the MIT license.
Contents
Using arelay.app
Self-hosting
Reference
Using arelay.app
The hosted service at arelay.app handles deployment, database, storage, backups, TLS, and updates. The security model is the same as self-hosting: encrypted deliveries are decrypted in your browser, not on the server. The one exception is approving an email draft: the decrypted email is sent to the server in that request so it can be delivered, and is never stored as plaintext. Email itself is not end-to-end encrypted, so the recipient's mail provider sees the same content permanently.
Get started
- Open arelay.app and create an account with a passkey.
- Accept the Terms of Service and Privacy Policy during signup (required on the hosted service).
- Complete Set up encryption on first portal visit (passkey + recovery key).
- In the portal, open Account → Agent tokens and create a named token for each agent or integration.
- Copy the token once — it is shown only at creation time.
- Your inbox updates automatically when agents send deliveries (refresh every few seconds).
Sign-in uses passkeys (WebAuthn), not passwords or social login. No Google, Apple, or Microsoft account is required. If terms or privacy are updated later, existing accounts are prompted to accept the new versions before continuing.
Connect your AI agent
Agent Relay is agent-agnostic: any client that speaks HTTP and follows the E2EE envelope format can deliver. Pick the path that fits your setup:
Whichever path you choose, set these where your agent runs — never commit tokens:
Every request uses Authorization: Bearer <ARELAY_TOKEN>.
CLI & MCP server
The @arelay/cli npm package wraps the API and all envelope encryption.
Deliver from any shell:
Or register it as an MCP server so agents can deliver work natively
(deliver_to_inbox, list_inbox_sessions, submit_email_draft tools):
npx -y @arelay/cli check verifies the token and encryption setup. The package also exports a
typed SDK (import { ArelayClient } from '@arelay/cli') — see its
README. Self-hosters point it at their deployment with
ARELAY_URL.
Agent skill
Install the agent-relay skill (requires
portal E2EE setup):
Hermes Agent:
Direct HTTP API
Anything that can POST JSON can deliver: fetch the public key from
GET /api/agent/e2ee/config, encrypt locally, create a session, and upload artifacts. See
the API reference
and the bundled Node reference scripts in the skill for working encryption code.
Platform plugins
Hermes cron (--deliver arelay) needs arelay-hermes-plugin.
Cron runs in the gateway, not the interactive CLI — put credentials in
~/.hermes/.env (including AGENT_RELAY_HOME_CHANNEL=https://arelay.app), install the
plugin, then restart the gateway:
Encryption (required)
All agent deliveries must be end-to-end encrypted. Complete Set up encryption during
Get started; unlock in the browser when viewing the inbox. Agents fetch your
public key from GET /api/agent/e2ee/config and upload encrypted sessions and artifacts.
Plaintext agent payloads return 400 (plaintext_not_allowed); unconfigured accounts get
428 (e2ee_required).
Email Review Relay
On arelay.app, agents POST encrypted email drafts; each appears as an inbox session you Approve (send via your Cloudflare credentials) or Reject.
- Account → Email sending (Cloudflare API) — save your Cloudflare Account ID and API token (encrypted server-side; used only when you approve).
- Use the agent-relay skill — drafts use the same E2EE envelopes as file deliveries.
API details and self-host setup: Plugins.
Features
- Mobile-friendly email-style inbox
- Email Review Relay: agents draft outbound mail; you approve before send
- Artifact preview and download (text, Markdown, HTML, PDF, images)
- Sandboxed HTML/Markdown preview (external links and media stripped)
- End-to-end encrypted deliveries only
- Passkey sign-in (one passkey per account)
- Named agent API tokens with per-token revoke
- Storage limits: 25 MB per artifact, 500 MB per account
Self-hosting
Run Agent Relay on your own PostgreSQL and S3-compatible storage. Useful when you need full data residency control or a private deployment.
Development setup
Open http://127.0.0.1:5173 and create an account with a passkey. Without Cloudflare or SMTP configured, verification codes are printed to the server console.
Database schema changes are managed with Drizzle migrations. Edit the Drizzle schema,
generate a migration with npm run db:generate, review the generated SQL, then apply it
with npm run db:migrate:local.
Run unit tests:
To test migrations against a disposable empty database, set TEST_DATABASE_URL and run:
Environment variables
When both Cloudflare Email Sending and SMTP are configured, Cloudflare is used.
Deploy to production
Requirements:
- PostgreSQL
- S3-compatible object storage (see
scripts/iam-agent-relay-s3-policy.jsonfor a minimal AWS IAM policy scoped to theagent-relay/prefix) - Email delivery for account verification (Cloudflare Email Sending or SMTP)
- A stable HTTPS origin with correct WebAuthn RP settings
Build and run:
For a fresh production database, run migrations before starting the app. If you are moving
early account/auth records from an old database, copy only users, webauthn_credentials,
e2ee_config, and agent_api_tokens; leave inbox/content/plugin tables empty.
Railway:
- Create a new service from github.com/mmmikael/arelay.
- Add PostgreSQL and link
DATABASE_URL. - Add S3-compatible storage credentials.
- Set
SESSION_SECRET. - Set
WEBAUTHN_RP_IDandWEBAUTHN_ORIGINfor your domain. - Configure email delivery: Cloudflare Email Sending (
CLOUDFLARE_ACCOUNT_ID,CLOUDFLARE_API_TOKEN,EMAIL_FROM) or SMTP. - Set
NODE_ENV=production. - Set
BODY_SIZE_LIMIT=140M(or rely onnpm start, which defaults to140Mwhen unset). A value pinned to the old40Mrejects the 100 MB artifacts the Pro plan allows. - Migrations run automatically on each deploy via
railway.toml(preDeployCommand: npm run db:migrate). - Use
npm run buildas the build command andnpm startas the start command.
Existing production database (legacy ensureSchema → Drizzle): on the first Drizzle deploy,
npm run db:migrate detects an existing users table, records the baseline migration in
drizzle.__drizzle_migrations without re-running CREATE TABLE, then applies any newer
migrations. No manual cutover is required for arelay.app-style databases that already match
the current schema. For a truly fresh host, run migrations on an empty database instead; if
you are rebuilding from scratch, copy only users, webauthn_credentials, e2ee_config, and
agent_api_tokens from the old database.
Point agents at your deployment URL: ARELAY_URL=https://your-domain.example.
Backup and restore (self-hosting)
Agent Relay is not a backup or archival service — keep independent copies of important content. Self-hosted operators are responsible for backups (see SECURITY.md).
Back up both stores; restoring only one leaves the deployment broken:
Backup
Use your host's managed snapshots (Railway Postgres, S3 versioning, etc.) when available.
Restore
- Restore PostgreSQL, then object storage under the same
S3_BUCKETandS3_PREFIX. - Keep the same
SESSION_SECRET— encrypted Cloudflare credentials in the database cannot be decrypted after restore if this value changes. - Run
npm run db:migratebefore starting the app (applies any migrations newer than the dump). - Spot-check: sign in with a passkey, open a session, download an artifact.
Delivered content stays E2EE-encrypted in backups. Recovery still requires the account passkey and portal unlock; backups do not grant server-side plaintext access.
Plugins
Optional features ship as plugins so minimal self-hosts stay lean. Plugin tables are included in the Drizzle migrations; enable each plugin with an environment variable to turn on its APIs and UI.
Email Review Relay
Agents submit outbound email drafts for human review before send. On arelay.app
the plugin is enabled by default; self-hosters set EMAIL_REVIEW_RELAY_ENABLED=true. When
disabled, plugin APIs return 404 and the portal UI is hidden.
Account setup (portal): open Account (/portal/account) → Email sending (Cloudflare API)
and paste your Cloudflare Account ID and API token.
These per-user credentials are encrypted server-side and used only when you approve a draft.
System CLOUDFLARE_* env vars remain for signup verification only.
Agent API: authenticated agents POST encrypted drafts to /api/agent/email-drafts
(encrypted: true with envelope fields). Plaintext drafts are rejected.
The response includes the inbox session and draft (status: pending). Poll draft status
with GET /api/agent/email-drafts/{id} or GET /api/agent/sessions/{id} (includes
email_draft when applicable).
Human review (portal): open the session, decrypt and preview the HTML, then Approve (sends via your Cloudflare credentials) or Reject (no send). Approve sends decrypted fields in the request body so mail can be sent without storing plaintext server-side. Approve requires Cloudflare Email Sending to be configured on the account.
Base64 data:image/... sources in approved HTML are converted in the browser to CID inline
attachments before sanitization and send. This keeps the encrypted portal preview
self-contained while making embedded images compatible with Gmail and other email clients.
Tech stack
- SvelteKit 2, Svelte 5, TypeScript, Tailwind CSS
- PostgreSQL, S3-compatible object storage
- WebAuthn/passkeys
Contributing & help wanted
Agent Relay is open source (MIT) and built in the open — contributions of all sizes are welcome, and there's real room to have an impact while the project is young.
The fastest way in is to build an integration. Most integrations talk to the HTTP API
through the @arelay/cli SDK and never touch the core app, so you can
ship something useful without learning the whole codebase or the encryption internals. The
Building integrations guide walks through four levels
— from a dozen-line delivery script to a full server-side feature plugin — with working
examples.
A few things that would genuinely help:
- Delivery bridges and notifiers (Slack, Discord, Telegram)
- A GitHub Action for delivering CI artifacts
- Platform plugins for more agent runners
- An additional email backend (e.g. SMTP) for Email Review Relay
See CONTRIBUTING.md for dev setup and conventions. Rough ideas are welcome — open an issue to start a conversation.
Security model
The hosted service publishes a plain-language security page: what is encrypted, what the server can see, the one exception at approval time, the credentials it holds, subprocessors, and known limits. In brief:
- Human login uses passkeys and signed session cookies.
- Agent access uses named bearer tokens; only hashes are stored server-side.
- Agent tokens require E2EE setup; an encrypted copy of each token is stored for reveal in the browser.
- All agent content uses P-256 ECDH and AES-256-GCM envelopes; decryption happens client-side.
- Server preview/download/archive APIs return
e2ee_only; the portal decrypts via ciphertext endpoints. - HTML and Markdown previews render in sandboxed iframes; external URLs are stripped. Users can optionally open HTML artifacts or email drafts in a new browser tab to view unsanitized content they trust.
- Artifact uploads are capped at 25 MB per file and 500 MB per account.
- Download and preview URLs are short-lived.
Report security issues privately as described in SECURITY.md.
License
Agent Relay is released under the MIT License.
Source: README.md at commit fef67ad
Tools
0Version history
1- v0.2.0LatestSep 16, 2026
Score
50% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
- Reached in the last 30 days25 pts
Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.
- Auth declared10 pts
Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.
- Tools listed15 pts
Answer tools/list without a credential, or ship a package that starts in a sandbox without one.
- DescriptionRequired15 pts
- Logo10 pts
- Documentation15 pts
- License10 pts