See every hidden agent your orchestrator runs.

The leanest way to orchestrate: Claude starts small Codex workers in hidden chats, gives each only what it needs, and replaces them once their context grows. Codex never shows those chats. Murmuratio does: whether a worker is still running, which model it got, how full its context is, and how fast you spend your limits. Made for control freaks and anyone with a touch of OCD.

  • macOS
  • Python 3.11, dependency-free
  • local only
  • telemetry-free
  • Claude Code
  • Codex
  • MIT license
live demo, sample data

Agents

4 running1 needs you3 done today

Claude limitsread from the Claude Code status lineWindowUsedResets5 hours20%in 2hWeek35%in 3dOuter ring: share of the limit usedInner ring: time gone in this window20%of the Claude 5-hour limit usedResetsin 2hWindow gone3 of 5 hoursPacebelow elapsed timeOuter ring: share of the limit usedInner ring: time gone in this window5h2h35%of the Claude weekly limit usedResetsin 3dWindow gone4 of 7 daysPacebelow elapsed timeOuter ring: share of the limit usedInner ring: time gone in this windowweek3dCodex limitsread from your Codex accountWindowUsedResetsWeek52%in 2dOuter ring: share of the limit usedInner ring: time gone in this window52%of the Codex weekly limit usedResetsin 2dWindow gone5 of 7 daysPacebelow elapsed timeOuter ring: share of the limit usedInner ring: time gone in this windowweek2d

Atlas1 needs you2 running1 done

OrchestratorOpus 5.5

Review implementation

atlasOpus 5.5

Review implementation

atlasOpus 5.5

Review implementation

Finished 1

atlasOpus 5.5

Finish the dashboard

Northstar2 running

OrchestratorOpus 5.5

Review implementation

northstarGPT-6.1 Sol

Run the test suite

Harbor2 done

OrchestratorOpus 5.5

Review implementation

Finished 2

harborOpus 5.5

Update the settings

harborOpus 5.5

Review implementation

  1. this tool gets the code today
  2. outer ring: limit used inner ring: time gone in the window
  3. your orchestrator
  4. hidden codex worker: its model and what it is doing now
  5. context filling up? start a fresh worker
  6. waiting for you
  7. done workers sink here
Sample data.

Why it exists

Claude and Codex are good at different things: design, code, reviews. Some people pay for both and let one of them run the work. That one is the orchestrator.

If Claude is the orchestrator, the cheapest way to work is not to keep long chats open. Claude starts small Codex workers on its own, gives each one only what its task needs, and starts a new one when a worker's context gets too full. Less context, fewer tokens burned.

The catch: Codex does not show these workers. Claude says the work is running, but you cannot see how many workers it started, which model each one got or how full their context is.

Murmuratio shows that. It reads the local Claude Code and Codex files on your Mac and lists every agent by project, with workers under the orchestrator that started them: task, model, status, context use, and your Claude and Codex limits. Nothing leaves your machine.

How it works

Murmuratio is a small web page that runs on your Mac. It reads the files Claude Code and Codex already keep: Claude sessions in ~/.claude/projects, Codex chats in ~/.codex.

Every few seconds it reads them again and matches each Codex worker to the Claude session that started it. Workers appear under their orchestrator, grouped by project.

For every agent it shows the model, what it is doing now, how long it has run and how much of its context window is used. Claude limits come from the Claude Code status line, Codex limits from its local state.

What it does not do

It does not start, stop or steer your agents. It only reads.

It does not send anything anywhere. No account, no telemetry, no cloud. The page is served on 127.0.0.1 and is gone when you stop it.

RequirementsPython 3.11+; run commands from the repository.
Install./murmuratio install
Run./murmuratio run, then open 127.0.0.1:8765 in a browser
Uninstall./murmuratio uninstall