murmur

A Herdr plugin that shows your running AI coding agents as a live org chart. It shows who spawned whom, what each agent was told, what it has cost by model, what it carried into context, and which agents are waiting on you. You can answer or interrupt them without leaving the tab.

Herdr shows you terminals. murmur shows you the shape of the work.

⛔ 1 NEED YOU · w4 Bash git push ──────── opus-5.5 $2.55 · sonnet-5 $0.47  09:41
                                      STATUS        MODEL      TOKENS  COST │ ◆ SUBAGENT  general-purpose
▣ w4  infra                           ⛔ BLOCKED 4m               41k $0.32 │ "Fix lint gate"
└─ ◉ claude  main                     ⛔ wait 4m    sonnet-5  41k 20% $0.32 │ spawned by claude · release-prep
   └─ waiting: Bash git push                                                │ model      claude-sonnet-5 · Anthropic
▣ w2  webapp                          ● working                  186k $2.71 │ definition built-in (Claude Code)
└─ ◉ claude  release-prep             ● working     opus-5.5 118k 59% $2.71 │ CONTEXT LOAD  ~48k tok before any work
   ├─ ◆ general-purpose "Fix lint…"   ▸ Bash        sonnet-5     14k $0.06 │ ████████████████████████████████
   ├─ ◆ code-reviewer "Review diff"   ▸ Read        opus-5.5     31k $0.71 │ INSTRUCTION FILES (2) ───────────
   └─ ✓ 1 done  →                                                   $0.09 │ User     ~/.claude/CLAUDE.md  ~2k tok

• Runs locally and costs nothing. murmur never calls a model or the network. It reads state already on your disk: Herdr's API and Claude Code's own session transcripts.
• Pure Node, no dependencies. Nothing to npm install and no build step.
• Works on Windows, macOS and Linux.

Requirements

• Herdr 0.7.0 or newer
• Node.js 18 or newer on your PATH
• Claude Code running in your Herdr panes. Other agents appear in the tree, but only Claude Code writes the transcripts murmur reads usage from.

Installation

From GitHub (recommended)

herdr plugin install AFetisa/herdr-murmur

To pin a release or branch, add --ref <tag-or-branch>.

From a local clone (for hacking on it)

git clone https://github.com/AFetisa/herdr-murmur
herdr plugin link /absolute/path/to/herdr-murmur

A linked plugin runs straight from your clone. Edits take effect the next time you open the pane.

Check it installed

herdr plugin list
# portablelabs.murmur (murmur) enabled

Open it

Open murmur from Herdr's plugin menu, or run:

herdr plugin pane open --plugin portablelabs.murmur --entrypoint murmur --placement tab

Update and uninstall

herdr plugin uninstall portablelabs.murmur && herdr plugin install AFetisa/herdr-murmur   # update to the latest version
herdr plugin disable portablelabs.murmur                  # turn it off, keep it installed
herdr plugin uninstall portablelabs.murmur                # remove (use `unlink` for a linked clone)

If you change herdr-plugin.toml in a linked clone, run unlink then link again. Herdr caches the manifest when the plugin is linked.

Using it

Key	Action
↑ ↓	Select a project, agent or subagent
⏎	Jump to that agent's pane in Herdr
y / n	Approve / deny a blocked agent's prompt (asks you to confirm)
x	Interrupt a working agent (asks you to confirm)
→ ←	Expand / collapse finished subagents
i / t	Briefing panel / tokens-by-model panel
[ ]	Scroll the panel
c	Context view: instruction files and hooks across all agents
s	Skill lens: skills, subagent types, MCP servers and commands, ranked
u	Inventory: installed plugins, MCP servers and connectors, and whether they are used
r / q	Rescan / quit

At 120 columns or wider, the briefing panel sits beside the tree and follows your selection.

Features

Org chart	Project → agent → subagent → sub-subagent, with real spawn depth and aligned status, model, token and cost columns
Waiting on you first	Blocked agents sort to the top with how long they have waited, and the header says how many need you
Cost by provider and model	Every turn is priced at its own model's rate, then rolled up per subagent, session, project and in total. The header shows the model mix
Briefing	For any row: who spawned it, its model, the .md file that defines it, the instruction files loaded into its context (from Claude Code's own record), skills available and used, health warnings, cost by model and the full task prompt
Context load	A stacked bar of what an agent carried before doing any work: tool schemas, instruction files, skill list, agent list, MCP instructions, deferred tools, hook-injected text and the task
Context view (c)	The reverse lookup across all agents. Each instruction file shows how many agents loaded it and its total tokens; each hook shows runs, time, injected text, errors and timeouts
Answer and interrupt	A blocked agent's briefing shows its actual screen, so you see the exact prompt. y, n and x act on it after a second confirming y
Health verdicts	thrashing, flaky, looping, expensive, sprawl, compacting and starved, each showing the number that triggered it
Stall detector	An agent marked working with no transcript writes for 90 s shows as ⚠ stalled
Skill lens and inventory	What you actually use, and what you carry for nothing
ASCII mode	MURMUR_ASCII=1 for terminals without Unicode; NO_COLOR=1 for monochrome

Privacy and safety

murmur is a local viewer. Here is exactly what it touches.

It reads:

• Herdr's local API: the pane list, statuses and titles.
• Claude Code's transcripts and subagent metadata under ~/.claude/projects/.
• Instruction files and agent definitions under your project and ~/.claude/.
• The visible screen of a pane, but only when that agent is waiting on a prompt.

It never:

• makes network requests or calls a model
• writes to your projects, transcripts or Claude Code settings
• sends keys to an agent unless you press y, n or x and then confirm with y

Keys it can send, through Herdr's own herdr pane send-keys:

Action	Sends	When
Approve	1	Only while that agent is showing a prompt
Deny	Esc	Only while that agent is showing a prompt
Interrupt	Esc	Only while that agent is working

The agent's status is checked again at the moment of sending. If it has moved on, nothing is sent.

Untrusted text is sanitised. Transcripts and screens can contain text from anywhere, such as web pages an agent fetched. murmur strips every terminal escape sequence and control character except its own colour codes before drawing. Transcript content cannot move your cursor, rename your window or write to your clipboard.

Watch what you share. The briefing shows full task prompts and file paths. Treat a screenshot or screen share of murmur like one of your agent's terminal.

Its only file write is a small inventory cache in Herdr's plugin state directory, or your OS temp directory when running outside Herdr.

Configuration

File / variable	Purpose
pricing.json	$ per million tokens and context window per model family. Update it for your rates and plan
thresholds.json	Health-verdict thresholds
MURMUR_ASCII=1	ASCII-only drawing
NO_COLOR=1	Monochrome output
HERDR_BIN_PATH	Path to the herdr binary if it is not on your PATH

What the numbers mean

• Token counts and subagent costs are measured from Claude Code's usage records.
• Context-load sizes are estimates at 4 characters per token, shown with a ~.
• Skill, MCP and slash-command costs are estimates, also shown with a ~. They run inside an agent, so they get a share of that agent's cost in proportion to their tool calls. murmur never raises a health flag from an estimate.
• Inventory verdicts are conservative, because acting on them means removing things:
  • never used needs a zero count and no last-used timestamp.
  • orphaned needs positive evidence that the plugin is gone.
  • Rows with ambiguous evidence are dropped rather than guessed at.

Troubleshooting

Symptom	Fix
The tab opens and closes at once	Run node src/app.js --once in the plugin folder to see the error. Check herdr plugin log list
Tab crashes on Windows with EISDIR: lstat 'C:'	Fixed in the manifest (--preserve-symlinks flags). Update the plugin, or unlink and link again after pulling
Status lags by about a second on Windows	Expected. Windows polls every 1.5 s (see below)
Two panes show the same cost	They share a folder and have near-identical titles; see "Pane-to-session matching" below
Garbled boxes	MURMUR_ASCII=1
Costs look wrong	Update pricing.json for your rates

Known limits

• Herdr can miss a permission prompt. Herdr 0.7.5 reports a pane sitting on a Claude Code permission dialog as done. murmur therefore treats an agent as waiting when its transcript has an unanswered tool call and its screen shows a Claude Code prompt.
• No live push events on Windows. Node cannot connect to Herdr's AF_UNIX socket on Windows, so murmur polls herdr api snapshot every 1.5 s. macOS and Linux get live events. This affects latency only.
• Pane-to-session matching is by title. Herdr 0.7.5 does not say which session a pane runs. murmur matches the pane title against the session's own title and falls back to the newest unclaimed session. Herdr 0.8's agent_session field will replace this.

Tested against Herdr 0.7.5-preview and Claude Code transcripts as of 2026-09. Fields that only exist in newer Herdr versions are optional, so murmur picks them up automatically when you upgrade.

Development

npm test                 # every module's built-in self-test; no framework, no dependencies
node src/app.js          # run standalone in any terminal
node src/app.js --once   # print one frame and exit
node src/render.js       # render the built-in demo fixture

Module	Role
src/herdr.js	Herdr API and CLI
src/claude.js	Transcript parsing: usage by model, instruction files, context load, hooks
src/render.js	Tree and panels
src/context.js	Context view
src/skills.js and src/lens.js	Skill profiling and the skill lens
src/inventory.js	Installed-versus-used inventory
src/app.js	State, keys and actions

Licence

Apache-2.0.