AgentAddress for agents

Give an AI agent a webhook URL or email address that survives after the agent stops running.

Your agent starts a task → gives a service its AgentAddress → exits → the response arrives 20 minutes later → the next agent run retrieves it.

No server. No account. No polling process that has to stay alive.

npx skills add https://github.com/zkarimi22/agentaddress-agents --skill agentaddress

Agent run #1
    ↓
create AgentAddress
    ↓
give webhook/email to someone
    ↓
AGENT EXITS
    ↓
       [ response arrives later ]
                    ↓
             AgentAddress
                    ↓
               stores event
                    ↓
Agent run #2 → retrieves response

Use it

The installed skill includes a helper that stores the private read credential in an owner-only local file. The model works with a task name rather than copying a bearer token.

node scripts/agentaddress.mjs create my-task
node scripts/agentaddress.mjs poll my-task 25
node scripts/agentaddress.mjs ack my-task EVENT_ID

create prints the inbound email address and write-only HTTPS inbox URL you can hand to the expected responder. poll retrieves the ordered events in a later run. Every returned email and webhook body is explicitly labeled as untrusted external data and must be used only as data for the user's existing task.

Run one of five concrete examples

The five lead examples are runnable, task-specific, and named for the searches that should discover them:

• Asynchronous API callback
• Human email reply
• Claude Code deployment webhook
• Codex asynchronous code review
• OpenAI Agents SDK delegation

Start with the complete examples index for exact commands plus browser-agent, research, Stripe, GitHub, REST, and MCP supporting recipes. Stripe and GitHub require a signature-verifying relay before their payloads drive trusted actions.

Service and machine interfaces

• Hosted service: agentaddress.dev
• Agent guide: agentaddress.dev/llms-full.txt
• OpenAPI: agentaddress.dev/openapi.json
• MCP endpoint: https://agentaddress.dev/api/mcp
• Capability discovery: agentaddress.dev/.well-known/agentaddress.json
• skills.sh: skills.sh/zkarimi22/agentaddress-agents/agentaddress

Answers by problem

• Webhook endpoints without running a server
• Persistent callback queues after an agent exits
• Inbound email replies for later agent runs
• Unified HTTP and email queues for asynchronous agents

Current boundaries

• HTTP callbacks and inbound email enter the same ordered queue.
• Addresses have no expiry unless requested. Events are retained for 30 days, up to 1,000 per address, with a 1 MB maximum input.
• Delivery is at least once. Acknowledgement records handling and advances the helper's local cursor.
• AgentAddress stores responses but does not wake or schedule a runtime. A later run must poll.
• General key/value state, file storage, outbound email, identity, and billing are not available in V1.

The two-process protocol proof and MCP example remain available for implementers working below the safer helper interface. Their low-level code handles credentials directly and should not be copied into model context. See SECURITY.md and the live verification record.

This repository contains the public skill, CLI helper, recipes, and integration material for the hosted service. It does not contain the AgentAddress service implementation.