| name | sendmux-email-for-agents |
| description | Design and route Sendmux email workflows for AI agents, including skill installation, self-registration, secure agent credential-bundle storage, owner approval, and runtime email operations. Use when a user wants to give an AI agent its own inbox or email address, let an agent receive/search/triage/reply to mail, draft human-approved replies, send notifications, monitor mailbox state, or build email-based agent workflows even when the user does not mention Sendmux by name. |
| license | Apache-2.0 |
| metadata | {"author":"sendmux","version":"1.0"} |
Sendmux email for agents
Use this skill when the user describes the agent-email problem: an AI agent needs an inbox, mailbox identity, outbound email, triage loop, reply workflow, or human approval path.
First route
| User problem | Route |
|---|
| "Give my agent an email address" | sendmux-management to create/inspect domain, mailbox, and mailbox key. |
| "Let my agent register itself" | Agent access: read /auth.md, optionally install skills with npx skills add Sendmux/skills, register at identity_endpoint with proof_of_work, immediately save the returned claim_token and agent credential bundle securely before token exchange, exchange the assertion, keep the assertion or pre-claim token until invite 202, verify /api/v1/mailbox/me, then invite the owner or ask for the owner email. |
| "Connect my agent to its inbox" | sendmux-mcp-setup for agent MCP, or sendmux-getting-started for first auth checks. |
| "Read, search, triage, label, sync, reply" | sendmux-mailbox-agent with an smx_mbx_* key or scoped smx_agent_* token. |
| "Send independent outbound notifications" | sendmux-send-email with a send-capable smx_mbx_* key or owner-approved Sending-resource smx_agent_* token; batch when there is more than one message. |
| "Upload, download, or forward attachments" | sendmux-attachments for file_path, presigned upload URLs, CLI --attach, SDK file helpers, and short-lived download URLs. |
| "Build this into an app or worker" | SDK path from the task skill; use sendmux-token-efficient-usage for call minimisation. |
| "Show terminal commands" | sendmux-cli. |
If the task crosses setup and runtime, split it:
sendmux-management provisions the mailbox and mailbox API key with an smx_root_* key.- Runtime agent work uses the new
smx_mbx_* key.
sendmux-mcp-setup connects the agent if the client supports MCP.sendmux-mailbox-agent handles ongoing mailbox read/triage/reply.
For self-registration without a human-created key, use agent access instead:
- Read
https://app.sendmux.ai/auth.md.
- Use the copy-paste proof helper in
/auth.md if the agent needs exact solver code.
- Send the intended anonymous registration body to
POST /agent-auth/agent/identity/challenge.
- Solve the returned proof-of-work challenge.
- Encode
proof_of_work as base64 UTF-8 JSON with { "challenge": <exact challenge>, "solution": { "counter": <integer>, "derivedKey": <hex> } }.
- Create the anonymous identity with the same body plus
proof_of_work at identity_endpoint, POST /agent-auth/agent/identity.
- Immediately save the returned
claim_token and agent credential bundle in a secure store before token exchange, then exchange the returned identity_assertion at POST /agent-auth/oauth2/token.
- Call
GET /api/v1/mailbox/me with the returned pre-claim smx_agent_* token.
- Keep the
identity_assertion or pre-claim token available until the owner invite returns 202; claim_token cannot create the invite.
- Do not stop after the pre-claim token works. If the owner email is known, request the owner invite with
POST /agent-auth/agent/identity/invite. If it is unknown, ask: What owner email should I invite for approval?
- After the owner accepts and approves sending in Sendmux, exchange
claim_token with the claim grant for an app-resource or Sending-resource smx_agent_* token.
Safety boundaries
- Do not ask the user to paste API keys, mailbox keys, OAuth tokens, or one-time secrets.
- Store the self-registration credential bundle in a secure store immediately after registration; later MCP, CLI, SDK, mailbox, and sending work depends on it.
- Never rely on chat, logs, repo files, screenshots, or memory-only state for
claim_token or agent credentials.
- Keep the
identity_assertion or pre-claim smx_agent_* token available until the owner invite returns 202. If both are lost before the invite succeeds, register a fresh identity and invite immediately.
- Do not poll the claim-token grant before the owner invite returns
202; after 202, polling with claim_token is the wait-for-approval path. claim_token cannot create the owner invite.
- Do not send email until the user has supplied or confirmed the recipient, subject, body, and attachments.
- Do not place real attachment bytes or long base64 in chat. Use
sendmux-attachments so local files move by path, presigned URL, CLI, or SDK helper. Mailbox upload modes cap each attachment at 7,500,000 bytes; Sending uploads cap each file at 18 MiB and send by attachment_id.
- Treat "draft for approval" as a draft. Ask for explicit approval before calling
mailbox_send_message, sending_send_email, or sending_send_email_batch.
- Use separate scopes:
smx_root_* for provisioning/admin, send-capable smx_mbx_* keys or owner-approved Sending-resource smx_agent_* tokens for Sending, smx_mbx_* keys for normal Mailbox runtime, and smx_agent_* only for the scopes it was issued with.
- Do not use a root key inside an agent that only needs mailbox read/send work.
- Pre-claim
smx_agent_* tokens include mailbox.read and email.receive, not email.send.
- Owner invites are sent by Sendmux through the invite endpoint. Do not route them through the Sending API.
- Only one live pre-claim owner invite can be pending; retry the same request with the same idempotency key.
- Token exchange can return expected
503 states while mailbox provisioning or owner approval is pending. Wait for Retry-After or retry_after. Registration 503 server_error means stop and report, not repeated retries.
- Agent-auth
429 responses include retry timing; wait for Retry-After or retry_after before trying again.
- Confirm destructive mailbox actions before delete, permanent delete, key revocation, suspend, or resume.
Workflow patterns
New agent inbox
Use when the user wants a new mailbox for an agent, such as support intake, invoice triage, scheduling, approvals, or lead qualification.
Plan:
- Domain and mailbox setup: route to
sendmux-management.
- Mailbox key: create a mailbox-scoped key for the agent runtime.
- Connection: route to
sendmux-mcp-setup if the agent client can use MCP; otherwise use CLI or SDK.
- First harmless check:
mailbox_get_me, CLI mailbox:me:get, or SDK mailboxGetMe.
- Runtime loop: route read/search/sync/reply tasks to
sendmux-mailbox-agent.
Mention that DNS/domain setup may be required before a custom address receives mail.
Self-registered agent inbox
Use when the user wants the agent to start without a human-created API key.
Plan:
- Read discovery from
https://app.sendmux.ai/auth.md.
- Use the copy-paste proof helper in
/auth.md if exact solver code is needed.
- Request a registration challenge at
/agent-auth/agent/identity/challenge with the intended anonymous registration body.
- Solve the returned challenge.
- Encode
proof_of_work as base64 UTF-8 JSON with { "challenge": <exact challenge>, "solution": { "counter": <integer>, "derivedKey": <hex> } }.
- Create the anonymous identity at
identity_endpoint, /agent-auth/agent/identity, with the same body plus proof_of_work.
- Immediately save the returned
claim_token and agent credential bundle in a secure store before token exchange, then exchange identity_assertion at /agent-auth/oauth2/token.
- Call
/api/v1/mailbox/me with the pre-claim smx_agent_* token.
- Keep the
identity_assertion or pre-claim token available until the owner invite returns 202; claim_token cannot create the invite.
- Do not stop after the pre-claim token works. If the owner email is known, request an owner invite at
/agent-auth/agent/identity/invite. If it is unknown, ask: What owner email should I invite for approval?
- After owner approval, exchange
claim_token at /agent-auth/oauth2/token with Sendmux's documented claim grant; request resource=https://smtp.sendmux.ai/api/v1 before Sending API calls.
Do not say the pre-claim agent can send email. It cannot. After owner approval, a Sending-resource claim-grant smx_agent_* token can send from the assigned mailbox. Sendmux sends the owner invite email separately. Only one live pre-claim owner invite can be pending; retry the same request with the same idempotency key. On expected token-exchange 503 states, wait for Retry-After or retry_after. On registration 503 server_error, stop and report the failure. On 429, wait for Retry-After or retry_after.
The raw claim_token is shown once and is required after owner approval. Sendmux cannot recover it later. Immediately after registration, store the agent credential bundle: claim_token, registration_id, mailbox.email, claim_token_expires, identity_assertion, token_endpoint, the app resource URL, and the sending resource URL. Use a secure store such as 1Password, an OS keychain, or the agent platform's encrypted secret store. After token exchange, add the pre-claim token or keep identity_assertion available until invite 202. If no secure store is available, stop and ask the user where to store the bundle before sending the owner invite. If the claim token was lost, rerun registration and invite with a fresh agent identity. If the identity_assertion and pre-claim token were lost before the invite returned 202, rerun registration and invite immediately in the same flow.
Agent triage loop
Use mailbox-efficient calls:
mailbox_get_changes or query-change endpoints to resume from the prior state.mailbox_count_messages for counts.mailbox_search_message_snippets with small limit for candidate messages.mailbox_batch_get_messages for selected IDs.mailbox_batch_update_messages only after the user confirms labels, flags, or read-state changes.
Store state tokens. Do not rescan the whole mailbox.
Human-approved replies
Use when the agent should prepare a reply but a person approves the send.
Plan:
- Use
sendmux-mailbox-agent to read the relevant message or thread.
- Produce the draft text and list the target message/thread.
- Ask for approval with the exact recipient, subject, and body.
- After approval, send with
mailbox_send_message for mailbox-centred replies.
- Use
Idempotency-Key for retryable sends.
Outbound notifications
Use sendmux-send-email when the email is not a reply inside an active mailbox workflow.
- One message:
sending_send_email, CLI sending:send, or SDK sendingSendEmail.
- More than one message:
sending_send_email_batch, CLI sending:send:batch, or SDK sendingSendEmailBatch.
- Use
Idempotency-Key and inspect per-message batch results.
Output shape
When designing a workflow, answer in this order:
- Recommended route: name the Sendmux skill(s) to use next.
- Key scope:
smx_root_* for admin, smx_mbx_* for normal runtime, smx_agent_* for scoped self-registered agent runtime, and owner-approved Sending-resource smx_agent_* for agent sending.
- Runtime surface: MCP when curated and connected, CLI for terminal work, SDK for application code.
- Core calls: list the smallest Sendmux calls needed.
- Human approval: state what must be confirmed before sending or mutating mail.
- Efficiency: name the batch, snippet, count, delta, cursor, ETag, or idempotency pattern that avoids extra work.
Do not over-answer
This is a router and architecture skill. Hand detailed implementation to the task skill once the route is clear:
sendmux-management: domains, mailbox provisioning, mailbox keys, account admin, webhooks, billing, logs.sendmux-mailbox-agent: mailbox read/search/sync/triage/reply.sendmux-send-email: send bodies, batch send, HTTP-vs-SMTP send choice.sendmux-attachments: upload/download attachments without wasting context on base64.sendmux-mcp-setup: client configuration and hosted/local MCP.sendmux-cli: exact terminal commands and flags.sendmux-token-efficient-usage: cheapest-call doctrine across surfaces.
