AgentLater

Let an AI agent make an HTTP request later, and check whether it happened.

AgentLater is a hosted Model Context Protocol (MCP) server for agents that need to schedule HTTP work, wait for a job, or retrieve results in a later session. It sends requests and records delivery outcomes; it does not run the agent itself.

What agents can do

Need	Tools
Send a request later or in the background	schedule_http, queue_http (optionally serialize related requests in a named queue)
Repeat a request or poll until a job finishes	create_cron, watch_http
Receive webhooks and read results in a later session	create_hook, read_inbox, ack_inbox
Inspect, stop or retry work	get_task, list_tasks, cancel_task, redrive_task
Use credentials or trigger an agent workflow when results arrive	set_secret, set_wake

See the full tool reference for arguments, limits and delivery semantics. A wake URL must belong to something that can start an agent run; inbox events alone do not resume a chat.

Connect an MCP client

Use a client that supports remote Streamable HTTP and MCP OAuth. Add a remote server or custom connector named agentlater with this URL:

https://gateway.agentlater.io/mcp

For clients using an mcpServers JSON configuration with HTTP transport, the configuration is:

{
  "mcpServers": {
    "agentlater": {
      "type": "http",
      "url": "https://gateway.agentlater.io/mcp"
    }
  }
}

Configuration formats vary by client; use its remote-server settings if it does not accept this shape. The endpoint and transport are also declared in server.json.

Authentication: complete the OAuth flow opened by your client and click Connect on the AgentLater consent page. There is no account to create or API key to paste into the configuration; the client manages OAuth tokens. Each approved connection owns a separate project and trial. Call create_project once after connecting (calling it again returns the same project): the trial includes 1,000 delivery attempts for 7 days. Use get_project to inspect the balance and get_payment_link when more attempts are needed.

For Claude Code, you can instead install the plugin, which includes the hosted server configuration and a skill for using the tools:

/plugin marketplace add thachp/agentlater
/plugin install agentlater@agentlater

See the plugin README for installation and migration details. Autonomous integrations can also use Agent Auth at the gateway; that identity and its project are separate from a hosted MCP connection.

A2A

An agent that speaks the Agent2Agent protocol (version 1.0, JSON-RPC) can schedule one-off deliveries and read, list and cancel them, signed in with Agent Auth and no human login. A2A is switched on per environment: it is on at https://gateway-staging.agentlater.io and may not be available on production yet, where its Agent Card answers 404 until it is.

# in a checkout, after npm ci
node apps/gateway/examples/a2a-client.mjs https://gateway-staging.agentlater.io

a2a-client.mjs is a complete client in one file: it reads the Agent Card, registers, makes its project, schedules a delivery and polls it to its end. Outside a checkout, copy the file and npm install @a2a-js/sdk@1.3.0 @auth/agent@0.6.2. It keeps its keys in ~/.agentlater/ (mode 0600). The auth sequence, payload schema, task states, errors and what is not offered through A2A (crons, watches, hooks, secrets, payments, streaming, push notifications, callbacks, lanes) are in the A2A docs and PRODUCT.md.

Example: trigger a background job through a webhook

A common use is calling your application's job endpoint after a delay—for example, generating an export in 15 minutes without keeping the agent session open.

After connecting, ask your agent:

Create my AgentLater project, then POST to my export webhook in 15 minutes with {"report": "daily-sales", "format": "csv"}. Use my stored webhook token, capture the response, and put the delivery result in my inbox.

The agent calls create_project with {}. If the endpoint requires a bearer token, first call set_secret with these arguments, replacing the placeholder with your endpoint's actual token and the host with your own:

{
  "name": "WEBHOOK_TOKEN",
  "value": "<your-webhook-token>",
  "allowed_hosts": ["api.example.com"]
}

Then call schedule_http with the following arguments. Replace the example URL with your own public HTTPS job endpoint. body is a JSON-encoded string, not an object.

{
  "url": "https://api.example.com/jobs/export",
  "method": "POST",
  "headers": {
    "Content-Type": "application/json",
    "Authorization": "Bearer {{secret:WEBHOOK_TOKEN}}"
  },
  "body": "{\"report\":\"daily-sales\",\"format\":\"csv\"}",
  "delay": "15m",
  "idempotency_key": "daily-sales-export-001",
  "capture": true,
  "callback_url": "inbox"
}

Use a new idempotency_key for each intended job; reusing the same key and request returns the original task. Your endpoint should also deduplicate deliveries using RunLater-Task-Id because delivery retries can reach it more than once. For an endpoint without authentication, omit the secret setup and the Authorization header.

Save the returned task_id and inspect it with get_task({"task_id": "<returned task_id>"}). In a later session, call get_project with {}, then read_inbox with {} if inbox_unread is above zero. After handling the events, call ack_inbox({"up_to": 1}), replacing 1 with the last seq you handled. These are MCP tool names and argument objects, not shell commands. No public callback URL is needed to receive the delivery report in the inbox. If the endpoint returns 202 Accepted, that confirms it accepted the job; use watch_http on its status URL when you also need to wait for the export to finish.

Delivery is at least once; succeeded means the destination returned HTTP 2xx. Destinations must be public HTTPS URLs, not localhost or private IP addresses. For authenticated requests, store the credential with set_secret and use {{secret:NAME}} in a header instead of pasting it into a request.

Limits and troubleshooting

• Schedule at most 7 days ahead. Cron expressions use UTC and run at most hourly; watches poll every 5 minutes to 24 hours. Captured responses keep the first 16 KiB of the body for 7 days after the task finishes.
• project_not_found: call create_project on the current connection before scheduling.
• budget_exhausted or trial_expired: inspect get_project and use get_payment_link for more attempts. Each delivery attempt, poll and URL callback attempt consumes budget; inbox delivery is free.
• Connection or sign-in fails: confirm the URL ends in /mcp and the client supports both Streamable HTTP and OAuth. A client limited to local stdio servers needs a supported bridge; see the Agent Auth bridge configuration.
• No result yet: inspect get_task with the returned task_id for its status and attempt history. Read the inbox using the same connection that scheduled the task; a new connection has a separate project.

Documentation and compatibility

AgentLater was formerly RunLater. Existing runlater plugin connections retain their own projects; installing the agentlater plugin creates a separate connection rather than moving existing work. Some internal package names and delivery headers (such as RunLater-Task-Id) retain the old name for compatibility.

The specs are PRODUCT.md (what it does), ARCHITECTURE.md (how it's built) and DESIGN.md (how it should feel, to agents and to people). The pre-rebuild implementation was removed at cutover; it is in git history before this change.

Layout

Path	What	Stack
apps/api	Private task API: ledger, scheduling, delivery	Rust · workers-rs
apps/gateway	Public entry for agents: Agent Auth, capabilities, MCP, A2A	TypeScript Worker
apps/web	agentlater.io landing, docs and payment page	Next.js · OpenNext
apps/receiver	Staging-only fixture the journey and the A2A run deliver to	TypeScript Worker
crates/core	Task domain, no Cloudflare deps	Rust
crates/egress	Destination policy and request signing	Rust
packages/contracts	Capability definitions, and the OpenAPI contract between gateway and API	TypeScript · OpenAPI 3.1
packages/tsconfig	Shared TypeScript config
plugins/agentlater	Claude Code plugin: the hosted MCP server and an agentlater skill	Markdown · JSON
plugins/runlater	The old plugin, deprecated and frozen so existing installs keep working	Markdown · JSON

Everything deploys to Cloudflare Workers, with staging and production environments in each app's wrangler.jsonc (apps/receiver has staging only); see docs/DEPLOY.md.

Develop

Requires Rust (pinned in rust-toolchain.toml), Node 22+, and cargo install worker-build --version 0.8.7 --locked.

npm install

cargo test && npm run typecheck && npm test

Every check CI runs, in one command (-- --fast skips the e2e suites):

npm run verify

The rules the code keeps, and the lints that enforce them, are in CODING_STANDARDS.md.

npm run dev -w @runlater/gateway

npm run dev -w @runlater/web

License

The code is MIT. Using the hosted service at agentlater.io is covered by its Terms of Service and Privacy Policy.