queue_add MCP tool

Sends a follow-up message to a running agent (empty slug = yourself). The message is queued and delivered when the agent is ready to process it, or, with delay_seconds/deliver_at, deferred: it parks in the agent’s durable workflow (surviving VM sleep, unlike an in-session timer) and delivers as a normal follow-up at the deadline, waking a sleeping VM. This is the reliable “resume me at T” primitive for autonomous agents; cancel an undelivered message with clear_queue using the returned follow-up ID. Returns an error if the agent is not running, for example after it completed or was killed. A sleeping agent wakes to take the message. Equivalent to murmur queue add in the CLI. Follow-ups are delivered according to the agent’s dequeue strategy: auto (the default: up to five of one kind per turn), all (drain everything), one (one per turn), or five (up to five per turn). Set front to deliver a message ahead of the queue, or preempt to make it the only work the agent runs next.

Parameters

Name	Type	Required	Description
`slug`	string	no	Agent task slug. Empty means self on VMs; with `delay_seconds`, that schedules your own resumption. Use an absolute path like `/w/{workspace}/{owner_provider}/{account}/{slug}` (e.g. `/w/backend/github_oauth/alice/fix-auth`) to target an agent outside your own path.
`description`	string	yes	Follow-up message to send to the running agent.
`tasks`	string[]	no	Checklist items to add to the agent’s task list.
`agent`	string	no	Agent persona override (e.g. `programmer`, `architect`).
`model`	string	no	Model override for this and subsequent turns (e.g. `claude-opus-4-8`). Sticky, like `agent`: empty leaves the agent’s current model unchanged. An explicit model overrides the persona’s model, exactly as at `spawn`. Must stay on the agent’s backend (claude vs codex); a cross-backend model is rejected.
`reasoning_effort`	string	no	Reasoning-effort override for this and subsequent turns. Sticky, like `model`: empty leaves the current effort unchanged. Claude: `low`, `medium`, `high`, `xhigh`, `max`. Codex: `none`, `low`, `medium`, `high`, `xhigh`.
`service_tier`	string	no	Codex service-tier override for this and subsequent turns: `standard`, `fast`, or `flex`. Sticky, like `model`: empty leaves the current tier unchanged. Codex agents only.
`preempt`	boolean	no	Interrupt the agent’s in-flight turn and discard pending queued follow-ups, then enqueue this one: atomically, so it becomes the only work the agent runs next. No interrupt happens when the agent is idle. Mutually exclusive with deferral and with `front`.
`front`	boolean	no	Deliver this message before everything already queued. Nothing is discarded and no turn is interrupted, so other queued work still runs, just after this. Combines with deferral: the placement applies when the deadline matures. Mutually exclusive with `preempt`.
`delay_seconds`	number	no	Defer delivery this many seconds from now (1 to 30 days’ worth of seconds). At most one of `delay_seconds` or `deliver_at`.
`deliver_at`	string	no	Defer delivery to this absolute RFC 3339 time (e.g. `2026-07-22T18:30:00Z`). Must be in the future and at most 30 days out. At most one of `delay_seconds` or `deliver_at`.
`attachments`	string[]	no	Local file paths to attach privately to this follow-up. Each path must name a regular file. Local `murmur mcp` server only: the hosted MCP endpoint does not expose it.
`workspace`	string	no	Workspace name. Overrides the session default. Required only when nothing else supplies one: off a VM, with no session default, and with a slug that is not a `/w/{workspace}/...` path.

Response

Returns a confirmation string carrying the message’s follow-up ID: follow-up <id> sent to <slug>, or for a deferral follow-up <id> deferred to <time> for <slug> (cancel with clear_queue + this ID). An empty slug renders as yourself.

Examples

Schedule your own resumption (from an agent VM)

{
  "delay_seconds": 2400,
  "description": "Deferred: check whether PR #123's CI finished"
}

The message parks in your workflow. Your VM can sleep in the meantime, and delivers as a normal follow-up in 40 minutes, waking the VM. A pending deferred message also holds a terminate-on-idle service agent alive until it delivers.

Send a follow-up message

{
  "slug": "fix-auth",
  "description": "Also fix the session expiry logic while you're in that file."
}

Response:

follow-up 4f6c1a2e-8d3b-4f0a-9c7e-2b5d8e1f3a4c sent to fix-auth

Send a follow-up with new tasks

{
  "slug": "fix-auth",
  "description": "Here are additional items to handle.",
  "tasks": ["Add rate limiting to the login endpoint", "Update the API docs"]
}

Interrupt then redirect

Use interrupt to stop the current turn, then send new instructions:

{
  "slug": "fix-auth",
  "description": "Stop what you're doing. Focus on the login handler instead."
}

Errors

Code	Meaning	What to do
`NOT_FOUND`	No agent with that slug exists, or the agent is not running.	Check the slug with `ls`. Use `spawn` with `resurrect: true` to resume a dead agent.
none	`delay_seconds and deliver_at are mutually exclusive`, `front and preempt are mutually exclusive`, or `delay_seconds must be between 1 and 2592000 (30 days)`.	Fix the arguments the message names.
none	`attachments are not supported on the hosted MCP (no local filesystem)`.	Attach files through the local `murmur mcp` server instead.
`UNAUTHENTICATED`	Identity token is missing or expired.	Check credentials.

murmur queue add: equivalent CLI command
interrupt: stop the current turn before sending a follow-up
clear_queue: discard queued and deferred follow-ups (all, or one by follow-up ID)
status: check agent state
spawn: create an agent

Getting Started

Core Concepts

Guides

queue_add MCP tool

Parameters

Response

Examples

Schedule your own resumption (from an agent VM)

Send a follow-up message

Send a follow-up with new tasks

Interrupt then redirect

Errors

​Parameters

​Response

​Examples

​Schedule your own resumption (from an agent VM)

​Send a follow-up message

​Send a follow-up with new tasks

​Interrupt then redirect

​Errors

​Related

Parameters

Response

Examples

Schedule your own resumption (from an agent VM)

Send a follow-up message

Send a follow-up with new tasks

Interrupt then redirect

Errors

Related