Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.murmur.dev/llms.txt

Use this file to discover all available pages before exploring further.

The Murmur dashboard is a web UI for managing your agent fleet. Use it to monitor running agents, steer ongoing work, spawn new agents, watch live sessions, and review the artifacts agents have produced (pull requests, commits, uploaded files, ports). Try it now: cloud.murmur.dev

Signing in

  1. Navigate to cloud.murmur.dev.
  2. Click Sign in with GitHub.
  3. You land on the dashboard, scoped to your organization’s tenant.
Sessions last 30 days and refresh automatically.
If you cannot sign in, check that the Macroscope GitHub App is installed on your GitHub organization and that your account is an org member.

The sidebar

Every agent you have access to shows up in the left sidebar. Each entry includes a colored status dot, the agent’s slug, and (if set) a one-line purpose. Click an entry to open its detail view.
Status dotWhat’s happening
Green (animated)Agent is starting (VM booting) or running
Yellow (animated)Agent is idle, sleeping, or waiting on checks
Red (animated)Agent failed or its VM is being destroyed
GrayAgent reached a terminal state (task complete or completed)
The sidebar collapses with Cmd/Ctrl+B if you want more room for the detail view.

Spawning from the dashboard

Click the + icon at the top of the sidebar to open the spawn modal. Fill in:
  • Task slug (optional): a short unique name used in URLs and logs (e.g. fix-flaky-test). Auto-generated from your prompt if left blank.
  • Prompt under “What should this agent do?”: plain English instructions
  • Workspace: which Workspace the agent runs in (resolves repos, environment, secrets)
  • Agent: which Agent Persona (e.g. programmer, tester, architect)
  • Mode: Autonomous (runs to completion), Interactive (you drive it via tmux), or Streaming (long-lived session)
  • Model: which Claude model powers the session
  • Identity (only shown when your tenant has Service Profiles configured): run as a bot identity instead of as yourself
Click Spawn. The agent shows up in the sidebar, the VM boots in 30 to 90 seconds, and the agent starts working immediately. This is the equivalent of murmur spawn from the CLI or mcp__murmur__spawn from a coding agent.

The agent detail view

Clicking an agent opens its detail view, which fills most of the dashboard. This is where you watch the agent work, send follow-ups, and inspect everything it produces. Across the top is the agent’s slug (click to copy). If you set a purpose at spawn time, it shows as a muted secondary line. Three toggle buttons on the right switch between views:
ToggleShortcutWhat you see
SessionAlt+SRaw streaming transcript of every message and tool call
VMAlt+VThe underlying VM: telemetry, logs, processes, files
DebugAlt+DInternal metadata: cost, model, workflow ID, branch links
Hold Option to see the keyboard shortcut overlay on every button.

Turns view (the default)

The main area is a scrolling thread of turns. Each turn is a discrete piece of work the agent did, packaged with what triggered it and what it produced.
  • Turn 1 is your initial prompt, rendered as Markdown with a copy button.
  • Subsequent turns each show a trigger card (a follow-up message from you, a PR comment, a PR review, a CI result, a file conflict on the base branch, a notify from a child agent, etc.) and what the agent did in response.
  • Cards collapse to a one-line summary (turn number, source icon or commenter avatar, label, preview text, relative time like “5m ago”). Click any card to expand it.
  • Expanded turns show the full input, the model badge (Claude or OpenAI icon plus model name), how long the turn ran, the agent’s written response, and a “View session” link that jumps into the raw session stream.
The currently running turn glows with an animated border so you can always spot what is in flight. Completed turns show their duration; errored turns are highlighted red. If the agent has set up a task checklist, it appears on the latest turn as a progress bar (e.g. “3 / 7”) with each item labeled Done, In progress, Pending, or Dormant.

Pending queue

If you send follow-ups (or events arrive) while the agent is busy, they stack into a single “N events queued” card below the last turn. Edit a queued message before it fires by clicking the pencil; trash the whole queue by clicking the trash icon (with confirmation).

Artifacts card

A small Artifacts card floats in the upper right of the content area, listing everything the agent has produced:
  • Pull requests with state icons, a “view diff” button that opens the inline diff viewer, and a “copy gh pr checkout” shortcut
  • Media and file uploads (thumbnails open in a lightbox)
  • Exposed VM ports with a ”+ Open Port” form to launch an ephemeral port-proxy URL
The card collapses to a single “Artifacts” pill (Alt+A) when you want it out of the way.

Follow-up composer

Anchored to the bottom is the follow-up composer: an auto-growing text box for the next instruction.
  • A persona picker to the left (Alt+P) lets you send the next message as a different persona, useful for “switch hats” moments.
  • The send button on the right becomes a stop button while the agent is running. One click sends a soft interrupt; a second click escalates to force-kill.
  • Messages you send while the agent is busy queue up in the pending queue (see above) and fire at the agent’s next natural break.

Session view

Toggle to Session (Alt+S) to see the raw stream: every message, tool call, and tool result with timestamps, cost markers, and PR callouts. Consecutive tool calls group into rows like “Edited 3 files, ran 2 commands” that expand on click. Cmd/Ctrl+F opens an in-viewer filter with a match count, handy when an agent has produced thousands of lines. If the agent runs in interactive or streaming mode, the composer at the bottom turns into a live chat input with its own send and interrupt controls.

VM panel

Toggle to VM (Alt+V) to inspect the agent’s virtual machine.
  • Sleep, Wake, and Kill action buttons live at the top.
  • Four collapsible panels below:
    • Telemetry: CPU, memory, and disk gauges with sparklines
    • Process: top processes by CPU and memory usage
    • Logs: live VM and journal logs with severity filter chips (DBG/INF/WRN/ERR) and a search box
    • Files: a tree-style live filesystem browser
Telemetry, Process, and Logs stay visible even when the VM is asleep (they’re backed by cached metrics). The file browser needs the VM awake; a “Wake VM” prompt appears if you click in while the VM is asleep.

Debug panel

Toggle to Debug (Alt+D) for everything structural about the agent run:
  • Phase, cost, duration, commit count
  • Full slug, short ID, session mode, completion check, on-idle behavior, dequeue strategy
  • Model, fast-mode flag, version
  • Links to the workspace, environment, branch, base branch, GitHub repo
  • VM instance name
  • Direct link to the Temporal workflow
  • Subscription list (branches and files the agent is watching)
  • Child agent list (clickable to navigate into)
Mostly useful when something is wrong and you need to file a bug or correlate with backend logs.

Diff viewer

When you click “view diff” on a PR in the Artifacts card, a full-height diff overlay slides over the right side of the page. It includes:
  • A file tree sidebar
  • Unified or split diff views
  • Inline PR review threads (reply to existing comments or compose new ones by clicking a line number)
Close with Esc or the X to return to the agent’s detail view.

Catalog browser

Browse your tenant’s configured resources (Workspaces, Placements, Environments, Images, Recipes, Roles, Flights, Service Profiles, Secrets) without leaving the dashboard. Each resource shows its current version, last author, and full configuration, with edit and history views for power users. Equivalent to murmur get from the CLI.

Pool status

A pool view shows your tenant’s VM pool health and capacity: total VMs, available VMs, assigned VMs, and configured limits. Useful for diagnosing slow agent starts or capacity issues. Equivalent to murmur pool status.

Keyboard shortcuts

ShortcutAction
Cmd/Ctrl+KOpen the command palette
Cmd/Ctrl+NNew agent
Cmd/Ctrl+BToggle sidebar
Cmd/Ctrl+DeleteKill the current agent (also accepts Backspace)
Alt+S / Alt+V / Alt+DSession / VM / Debug view
Alt+AToggle Artifacts card
Alt+POpen the agent picker in the composer
Space / bPage down / up in the active view
g / GJump to top / bottom
EscClose an open overlay (diff viewer, modal, spawn dialog)
Cmd/Ctrl+FFilter the session stream
The empty-state view lists these shortcuts alongside a “spawn your first agent” call to action.