AgentMicro
AgentMicro is a free, open-source macOS menu bar app that tracks the live state of your OpenAI Codex tasks. It reads local session data only, showing which tasks are thinking, waiting, finished or failing.
What is AgentMicro?
AgentMicro is a small macOS menu bar companion for developers who run several OpenAI Codex tasks at the same time. It answers a narrow but genuine question, which of my agents actually needs me right now, without making you reopen Codex and hunt through windows.
The menu bar icon is built from six tiles mirroring the first six tasks, and it animates while work is in progress. Opening the menu reveals the full queue. Each row carries a colour block, the project name, the task title and the duration of the current turn counted to the second; a lightning badge marks tasks running in Codex fast mode. Running tasks stay pinned at the top, followed by whatever changed most recently.
Five states are displayed: Idle in white when a task exists but has no current activity, Unread in green when a result is ready and unseen, Thinking in blue while Codex is processing, Needs input in orange when an approval or an answer is expected, and Error in red when the turn has failed. A sixth internal state, unknown, is deliberately shown as idle rather than introducing a misleading colour, because the project would rather admit ignorance than guess.
Clicking a Codex Desktop task opens the matching conversation. State refreshes from filesystem events, with a short safety poll running behind them.
The design is local-first and read-only. AgentMicro reads known Codex process and local session metadata and nothing else: it never uploads prompts, code, responses, command output or task history, does not read the Keychain and does not request Full Disk Access. Base mode needs no account and no Accessibility permission, while an optional enhanced detection mode exists, stays off by default, and never clicks, types or approves on your behalf.
It requires macOS 14 Sonoma or later plus Codex Desktop or Codex CLI, ships in 23 interface languages, and is free under the MIT licence. It is an independent, community-maintained project derived from Peter Steinberger's CodexBar, with no affiliation to OpenAI. Version 1 observes tasks and takes you back to them; it does not drive them.
What it does
- See the live state of every running Codex task from the macOS menu bar
- Tell at a glance which task is thinking, has an unread result, needs input, errored or is idle
- Reopen the matching Codex Desktop conversation with a single click
- Read the current turn duration, down to the second, for each task
- Spot tasks running in Codex fast mode through a lightning badge
- Keep active tasks pinned at the top of the queue, ahead of recently changed ones
- Switch between task-title, project-only and combined display modes
When to use AgentMicro / When not to
A quick filter to help you decide if AgentMicro is the right fit.
When to use AgentMicro
- Developers running several OpenAI Codex tasks in parallel and losing track of which one needs attention
- Codex Desktop and Codex CLI users working on macOS 14 Sonoma or later
- Privacy-conscious engineers who refuse to send prompts, code or task history to a third-party server
- Solo developers and indie builders who want a productivity gain without a subscription or an account
- Open-source enthusiasts who want to audit, fork or translate the tool they install
When not to use AgentMicro
- Windows and Linux users, since AgentMicro ships only for macOS 14 and above
- Teams working with a coding agent other than Codex, as version 1 is deliberately Codex-only
- Anyone expecting to act on tasks: the app never approves, stops, resumes or edits anything
- Users who need token or quota tracking, which is explicitly outside the product scope
- Organisations requiring a published privacy policy, terms of service and an identifiable legal vendor
How to use AgentMicro
A typical end-to-end flow, from setup to results.
- Check the prerequisites: macOS 14 Sonoma or later, with Codex Desktop or Codex CLI already producing local sessions
- Download the latest signed and notarised build from the project's GitHub Releases page, preferring the DMG when one is listed
- Open the DMG and drag AgentMicro.app onto the Applications folder, then launch it from /Applications
- Alternatively install the same build in one command through the official Homebrew tap: brew install --cask fizzy718/tap/agentmicro
- Or build it yourself with Xcode and Swift 6.2 or later, using make agentmicro-package
- Follow the first-launch guide, which walks through the five state colours
- Open Settings and choose how many recent tasks to display, anywhere from 1 to 20
- Pick a display mode: task title, project only, or the two combined
- Enable launch at login and recently completed task retention if you want the queue permanently available
- Turn on Enhanced Status Detection only if you want it, as it is off by default and only then requests Accessibility permission
Pros & Cons
Pros
- Completely free and open source under the MIT licence, so the code can be audited or forked
- Local-first and read-only: no prompt, code, response or task history ever leaves the Mac
- Works with no account and, in base mode, no Accessibility permission
- Installs in seconds from a signed, notarised DMG or a single Homebrew command
- Version 0.1.4 cut non-initial CPU samples from roughly 30-74% down to 0.7-4.2% on the same workload
- 23 interface languages, following the macOS system language by default
- Actively developed, with five releases published within the first eleven days
Cons
- macOS 14 or later only, with no Windows, Linux or mobile version
- Codex-only by design, as the other providers inherited from the CodexBar codebase are not connected
- Very young project: the repository was created on 29 July 2026 and sits at version 0.1.4, with no 1.0 yet
- No privacy policy, terms of service or legal notice, and no identifiable legal entity behind it
- Displayed state is derived from local metadata because Codex exposes no stable event API, so brief mismatches with the Codex interface can occur
- Purely observational: it cannot approve, stop, resume or edit a task, nor track token quotas
- No private support channel, with neither an email address nor a form, as everything goes through GitHub
Pricing & Plans
AgentMicro is offered entirely free of charge. There is no paid plan, no subscription and no in-app purchase, and consequently no lowest price point to report. The software is distributed under the MIT licence and its full source code is publicly available, the homepage stating plainly that the tool is free and open source. No free trial period exists either, the notion being inapplicable to a product that carries no paid tier whatsoever.
Data, GDPR & hosting
A consolidated view of how AgentMicro handles your data.
GDPR overview
There is no mention of the GDPR anywhere, neither on the website nor in the repository nor in the release notes. A targeted search across every collected page and the archived HTML returned zero occurrences of GDPR, data protection, DPA or subprocessor. No privacy policy, terms of service or legal notice is published, no Article 27 representative is designated and no data protection officer is named. This is an absence of documentation rather than a statement of non-compliance. Architecturally the exposure is narrow: the application runs locally, creates no account and sends no personal data to the developer, so there is little for a controller-processor framework to govern. A European organisation with formal vendor-assessment requirements will nonetheless find nothing here to assess.
Who owns the data?
AgentMicro publishes no terms of service and no privacy policy, so ownership is never defined contractually. In practice the question is largely moot, because the application transmits nothing. It reads local Codex process and session metadata on the user's own Mac and keeps the result there, so the developer receives no prompts, no source code, no responses, no command output and no task history, and therefore holds no user data whatsoever. The single outbound request documented is the AgentMicro update feed. Ownership consequently rests entirely with the user, established by architecture rather than by contract, which is a distinction worth keeping in mind.
Reuse rights
Because nothing leaves the machine, there is no vendor-side reuse to authorise or restrict. Users keep full and unrestricted use of their own Codex sessions, prompts and outputs, exactly as they would without the app installed. The software itself is distributed under the MIT licence, so its source code may be read, modified, redistributed and reused commercially provided the copyright notice is preserved, the original CodexBar notice included. No clause governs user content because no document exists to contain one. Here the absence of terms is itself the answer, and it happens to fall in the user's favour.
Data retention & training
Hosting summary
There is no user data hosting to describe, because AgentMicro stores nothing remotely. The application reads Codex process and session metadata from the user's own Mac and keeps everything there. No account exists, no server receives task data, and the developer publishes no jurisdiction because none applies to user content. The only outbound connection documented is the AgentMicro update feed, which requests release information rather than transmitting anything about the user's work. The marketing website is a separate matter and should not be confused with it. The domain agentmicro.cc resolves to 188.114.97.0, a Cloudflare anycast address geolocated in the United States, and it is registered through a Chinese registrar, Alibaba Cloud Computing Ltd. That concerns the single landing page only, not the software or the data it reads. No hosting country, no hosting region, no subprocessor list and no data processing agreement is published anywhere, which is coherent with an architecture that never collects anything in the first place.
Things to keep in mind
Risks and trade-offs to weigh before adopting AgentMicro.
- Constant visibility of a task queue can turn into a compulsion to check, since an always-present status light invites the same reflex loop as a notification badge
- Watching agents work is not the same as understanding what they produced, and a one-click return can encourage skimming rather than genuinely reviewing generated code
- Delegating more tasks in parallel simply because they are easy to track can quietly erode the depth of attention given to each one
- With no privacy policy, no terms of service and no identifiable legal entity, there is no contractual recourse should the software ever behave differently from its documentation
- Task titles stay readable to anyone looking at your screen, so on a shared display or during screen-sharing they can leak client names or project details, which is precisely why a project-only mode exists
- Enabling the optional Accessibility permission grants a broad level of system access and deserves a deliberate, informed decision rather than a reflex click during setup
- Displayed state can briefly diverge from what Codex actually shows, so acting on a colour alone, assuming a task has finished or failed, can mislead
Setup & Integrations
Technical difficulty
Setup is straightforward for anyone already using Codex. Download the signed, notarised DMG, drag the app into Applications and launch it, or install the same build with a single Homebrew command. No account is created, no API key is configured, and base mode requests no system permission at all. A first-launch guide explains the five state colours. The only technical route is building from source, which requires Xcode with Swift 6.2 or later. The real prerequisite is not the installation but the context: macOS 14 or later, with Codex Desktop or Codex CLI already producing local sessions.
Deployment
Integrations
Supported languages
Behind AgentMicro
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement AgentMicro.
Frequently asked questions
How much does AgentMicro cost?
What do I need before installing it?
Are my prompts or my source code sent anywhere?
Do I need to create an account?
Does it need Accessibility permission?
Can AgentMicro approve or stop a task for me?
Which task states does it show?
How do I install it?
Does it work with coding agents other than Codex?
What is its relationship with CodexBar?
Should you pick AgentMicro?
AgentMicro does one thing and states its limits plainly, which is rarer than it should be. If you run several Codex tasks at once on a Mac, it removes a genuine irritation: the constant window-switching required to find out which agent is still thinking and which one has been waiting on you for ten minutes. The five-state colour scheme is legible at a glance, and the refusal to invent a sixth state when the evidence is thin says something reassuring about how the rest is built.
The local-first, read-only design is the strongest argument in its favour. Nothing is uploaded, no account is created, and base mode does not even ask for Accessibility permission. For anyone whose task titles or prompts are confidential, that architecture is worth more than any privacy promise a cloud service could make.
The reservations concern maturity and paperwork rather than the software itself. The repository was created on 29 July 2026 and sits at version 0.1.4; there is no privacy policy, no terms of service, no legal notice and no identifiable company, only a GitHub account and an MIT licence. An organisation with a formal vendor-assessment process will have nothing to file. Displayed state is also derived from local metadata rather than from a stable Codex event API, so brief disagreements with the Codex interface remain possible, and the project acknowledges as much.
Being free and open source changes the calculation considerably: there is no lock-in, no billing relationship and no data to reclaim, and the code can be inspected before it is trusted. For a macOS developer already working with Codex, AgentMicro is an easy addition. For everyone else, it is simply not applicable.
- Choosing a selection results in a full page refresh.
- Opens in a new window.