macOS menu bar app · Claude Code · Codex · Qwen · Gemini · Cursor · +3 more

Agents at work,
watched from the side.

Run Claude Code with Codex and Qwen open beside it, and after a while you lose track of what each one is doing and who called whom. AgentBar sits in the menu bar and reconstructs it from the files your agents already write to disk: each agent's current state, the call graph between them, and a token-and-cost postmortem. An approval waiting on you is read from the last record of the raw log and pinned to the top. AgentBar itself spawns no CLI, no MCP server, no browser.

Get started Features

macOS 14+ · Swift 6 / SwiftUI · lives in the menu bar, opens a Graph window

FinderFileEditViewGoSat 8/1 10:33

AgentBar

ACTIVITY10s

Needs attention2

codexbar

Claude · sonnet-4.5 · 2m

Awaiting Bash approval (inferred from log tail)

Approval

12.4k

landing-page

Qwen · qwen3-coder · 6m

Input

8,204

Active1

agentbar

Codex · gpt-5-codex · 30s

Running

3,120

Recent1

docs-site

Cursor · 42m

Done

9,881

Manual·

agents watched: 8
processes spawned to read usage: 0
outbound requests by default: 0

Features

Watched from the side. Nothing spawned.

The popover and the Graph window show only records the agents wrote to disk themselves. Usage, history, and call relationships all rest on the same foundation.

POPOVER · Active tab + separate window

See the current state and the reasons side by side

Active

Lists each agent's current state (running / awaiting approval / waiting for input / done / failed), with pending approvals rising to the top. Watching runs on file polling. For Claude, a pending approval is detected from the trailing tool_use (uncompleted) and result records in the raw log. Not a guess.

Graph

Shows call relationships between agents as an indented tree. When Claude spawns codex exec and that Codex spawns claude -p, the chain is reconstructed from tool-call commands, timestamps, and working directories. Nodes are runs (execution units), so resuming never breaks the parent-child links.

Analysis

Reports tokens and estimated cost per provider, cache hit rate, the bottleneck (longest run), and the graph's max depth, and warns when the agent chain is too deep. Costs are estimates and may already be covered by a subscription. Not a basis for billing.

Rules

Lists global (~/.codex/AGENTS.md ~/.claude/CLAUDE.md, etc.) and project agent config files. Present files show in green, unconfigured ones in gray, and selecting one opens it for editing on the spot. The Compare toggle lines up global and project files so you can check for duplicate instructions line by line. Un-imported files and legacy config are flagged here too.

POPOVER · Usage / History tabs

Reads only files the agents already wrote

Usage and history never go through a CLI, MCP server, or browser. AgentBar reads files the agents are already writing.

Claude~/.claude/projects/**/*.jsonl+ rate-limits.json

Codex~/.codex/sessions/YYYY/MM/DD/*.jsonl

Qwen~/.qwen/projects/<project>/chats/*.jsonl

Gemini~/.gemini/tmp/<project>/chats/*.jsonl

Cursor…/Cursor/User/globalStorage/state.vscdbSQLite, read-only

Command Code—no local files · remote quota (opt-in)

OpenCode~/.local/share/opencode/storage/**/*.json

Hermes~/.hermes/state.dbSQLite, read-only

Usage shows daily tokens, quota (warning color from 80%), per-model breakdown, and estimated cost. History drills down from sessions in the last 60 days to conversation turns (user / assistant / tool).

Lightness is a feature, so it's proven at runtime

Updates and self-monitoring

Update

npm brew cask app binary — updates AI tools however they were installed. Runs only when you trigger it, with a hard timeout on every command.

Diag

Measures its own CPU time, memory (RSS), and files read, and reports them in the popover. The count of forbidden processes spawned is always 0, and it shows as a plain number.

ProcessGuard

Compares snapshots of descendant processes before and after each usage fetch, and raises an error if a forbidden process appeared. 'Spawns nothing' is implemented as a runtime check.

Verified by 172 tests, including regression tests for no-process-spawn and load boundaries.

Design

No agent is spawned to read usage

Usage, history, and the call relationships between agents can all be read from files the agents themselves wrote to disk. So AgentBar does not spawn agents to get usage. CodexBar spawns the CLI on a schedule to get the same data. The difference is design, not optimization.

What a single usage fetch looks like in each app.

One usage fetch in CodexBar

timer
claude CLI launches
MCP · OAuth · auto-update
browser · session created
Mac freezes

One usage fetch in AgentBar

timer
open JSONL
parse
display

Aspect	CodexBar	AgentBar
Where usage comes from	Spawns the agent CLI on a timer (a full Node.js stack with MCP, OAuth, and auto-update)	Reads JSONL / JSON files the agents already write to disk
Subprocesses	Spawned. Pulls in MCP servers, auto-update, even a browser	None
Network	CLI-dependent (49 GiB per day reported from auto-update alone)	No traffic by default. Only explicitly configured remote usage is an exception
On failure	Keeps retrying, discards the cache, loses data	Keeps the last good data and shows when it was fetched
Auth dialogs	Keychain prompt reappears every few minutes (multiple reports)	No auth dialogs in the background

The CodexBar column is based on reports in public issues. See docs/codexbar-issues.md for details.

From CodexBar's GitHub issues (excerpts)

200 issues surveyed, #1767–#2467 · as of 2026-07-27

#2052
Background probe triggers CLI auto-update
90.6 GiB downloaded over three days. 257–495 Claude CLI processes spawned per day.
#1844
Hourly delegated CLI refresh launches the default browser
A recovery path for misdetected token expiry called /usr/bin/open. During fast retries, a browser opened every 20 seconds to 3 minutes.
#2251
Each probe creates a new empty account session
36+ phantom sessions accumulated on claude.ai over a few weeks. Also tripped the provider's rate limits.
#1999
CodexBarCLI memory grew to 46 GB + 29 GB
Triggered a system-wide OOM (jetsam) on a 48 GB RAM machine. Similar leaks recurred from #1004.
#2241
CLI probe success rate: 8.6% (23 of 268)
Discarded the cached snapshot on failure, leaving gaps of up to 110 minutes.
#2306
Status item refresh spins MenuBarAgent forever
Menu bar froze for minutes at 100–125% CPU (macOS 27 beta).

Source: the public issues at github.com/steipete/CodexBar. The classification and summaries are collected in docs/codexbar-issues.md inside the repository.

Get started

Install the DMG, or build from source

The recommended install is the DMG from GitHub Releases (signed and notarized). Building from source works too. Once launched, it stays in both the menu bar and the Dock.

DMG (recommended)

Download from GitHub Releases. Signed with a Developer ID and notarized.

From source

git clone, then swift run. Building requires Xcode 16 / Swift 6.

Homebrew Cask

The cask definition is in the repository. The public tap is not published yet.

Download the DMG from Releases

macOS: 14 (Sonoma) or later. Apple Silicon and Intel.
Building from source: Xcode 16 / Swift 6 (Command Line Tools alone are not enough)
Dependencies: None (SwiftPM only)

Click the gauge icon to open a popover with five tabs: Active / Usage / History / Update / Diag. From the Active tab you can open the Graph window, which shows call relationships between agents as a tree. The popover layout is specified in docs/DESIGN.md in the repository.

terminal

# From the repository root
swift build

# Run. No window opens.
# A gauge icon appears in the menu bar.
swift run AgentBar

# Core logic tests (includes the no-process-spawn regression tests)
swift test

Agents at work,watched from the side.