← 开源
jazzenchen

agent-hud-open

A HUD for Claude Code, Codex and 25 more coding agents. Know your quotas, tokens and sessions at a glance, answer requests and jump back in one click. On Mac, Windows and Linux, your phone and your watch.

Rust
在 GitHub 打开
增长势头
+4124 小时新增 Star+6.9%
639
Star
90
Fork
11
贡献者
创建于 2026-10-22 · 更新于 2026-11-19 · 今日第 445 名
主要开发者
jazzenchenUltroncnnemo-fancy
README

Agent HUD

HUD for your agents.
On your screen. On your phone. On your wrist.

English · 简体中文 · Website · Discord · WeChat group · Supported agents · Documentation


Left: Notch mode at a MacBook's notch · Right: Simple mode on a display's left edge, with Separate agents on · Sample data

If Agent HUD is useful to you, a star helps other people find it.

Your agents work on your Mac. Agent HUD brings their progress to you. Check where they are at a glance, pick up a conversation, make the next decision: at your desk, and after you leave it.

Install

With Homebrew on a Mac:

brew install --cask jazzenchen/tap/agent-hud

With Node.js on a Mac, Windows or Linux (how it installs):

npx @agenthud/install

Or download the release. Either command installs the newest release, betas included, and on a Mac Agent HUD then updates itself. If it is already in /Applications, add --adopt to the Homebrew command.

New here? Start with Agents messaging each other and what works on each platform.

On your screen.

Your agents, at a glance. You write code, read docs or deal with something else while your agents keep working. The edge of your screen tells you who is still busy, who is done and who is waiting for you.

The details are close at hand, and what needs your approval is answered right there. Every agent's progress, gathered where one glance takes it in.

  • Glow: one colour per agent, from its most used quota: green under 70% used, yellow from 70%, red from 90%, with a warning before a quota runs out. It moves faster while an agent works.
  • Hover: quotas with their resets or burn rates, balances, the last 24 hours of tokens, and the sessions, those waiting for you first.
  • Requests: approve, deny or always allow, and answer an agent's questions, all on the HUD, the questions Claude asks in its desktop app included. Agent HUD never decides for you: a request you leave goes back to the agent after 10 minutes.
  • Jump back: one click takes you to a session's thread, tab or terminal pane.

Two modes. Each display runs one: Notch by default where there is a notch, Simple elsewhere.

Mode On screen
Notch At the notch, with the glow around it.
Simple A small floating panel on any screen edge: each agent's logo in a ring of its most used quota, with the percentage beside it. A logo becomes its bot while its agent works.

A display can also leave out its HUD, or only its messages, requests or paper planes: Settings → Display.

On your phone.

Keep moving when decisions are needed. Leave your computer and the work still goes with you. On iPhone, check progress, read replies and continue the conversation.

When an agent reaches something only you can decide, your phone shows you the context, and you answer the question or approve the next step. The task keeps going, and so does your day. iPhone Duo is supported.


Agent HUD for iPhone · Home and Sessions, and Stats on an unfolded iPhone Duo · Sample data

See Agent HUD Remote.

On your wrist.

Start from your wrist, continue on iPhone. An idea comes to you on a walk. Raise your wrist, say it, and it goes straight to the Mac at home. A thought caught in passing becomes a task in progress.

Read the latest reply on your watch. To discuss it or add detail, continue on iPhone. The idea starts on your wrist, and the conversation goes where you go.


A note spoken to Apple Watch becomes a Codex session on the Mac · Sample data

See Agent HUD Remote.

Built-in MCP.

Your agents use Agent HUD too. Its built-in MCP server gives them a channel to talk to each other, and the tools to build the UI you want.

Let your agents talk to each other.

When Claude Code finishes a change, it can ask a Codex session to review it and get Codex's answer back, or start a new session and hand the task over. Agent HUD already sees every session, so it carries messages across clients.

You can watch who is talking to whom: the sender folds its message into a paper plane and flies it to the receiver's mark, and the answer flies back the same way.

  • A message, not the conversation: only what the agent writes goes across; its session stays where it is. On one project, the project itself becomes the shared context, and the other agent reads the code for itself.
  • Twelve agents: Claude Code, Codex, Antigravity, DeepSeek Harness, Cursor, TRAE, WorkBuddy, ZCode, Qoder, Kimi Code, Pi and OpenCode can message each other, or start a new Claude Code, Codex, Antigravity or DeepSeek Harness session.
  • Queued by default: a message waits until its session is idle. Only when you ask for it can an agent steer one into the turn in progress of a Codex or DeepSeek Harness session.
  • Six are limited: Cursor, TRAE, WorkBuddy, ZCode, Qoder and Kimi Code take a message only when one of their turns ends, so an idle session gets it after you start another turn there. All but WorkBuddy also cap how many they take in a row.

How to turn it on and use it: Agents messaging each other.

Build your own UI.

Tell an agent how you want to see things: “Put Codex first and keep only the numbers I care about.”

Through MCP, the agent rearranges your Mac HUD and changes what it shows and how it looks. The change appears at once as a preview and is kept only when you press Save, so the HUD fits the way you work.

  • For Claude Code, Codex and any other MCP client.
  • An agent can read the HUD, propose a layout or a look for any display, and check the result with a screenshot in which accounts are masked.
  • The same tools answer questions about your usage, sessions and statistics, from the data the app already holds.

Worth keeping on

Know your token usage.

Know what your agents did, and how much they used. See your token use, look back at how it changed, and find where it went: which agents, models and sessions.

Quotas tell you how much is left. Token usage shows you how it was spent.



Statistics on the Mac · a year of usage, and one turn's sub-agents on a timeline · Sample data

Dimension What you can track
Time Today, the last 7, 30 or 90 days, or a year. Every token stays on your computer; the phone keeps 30 days.
Token kinds Input, output, reasoning, cache writes and cache reads, all together or any mix.
Agents and models Each agent's share, tokens and cost, and the models behind them.
Sessions Tokens with the sub-agents' part, cost, turns, calls, context fill and cache hits. Sub-agents count inside the session that started them.
Turns and calls The new tokens of every turn, and each call with its model, context and the tools it asked for.
Cache hits Today, the last 7 days and the last 30 days, each session, and each model within a session.
Cost An estimate at each vendor's list price, from its China or international price list, marked ≈. It is not a charge.
Plans What each plan's usage would cost at API list prices, against the plan's price.
Activity When your agents worked, by day over a year or by hour, in one heatmap.
Quotas How full each quota window got, day by day, and where it reset.
Balances DeepSeek's spend per day and the day its balance runs out.

See your agents come to life.

Familiar agent logos become small bots that act out thinking, working and waiting. With a face on every state, one glance tells you what is going on.

They give the HUD a character of its own: your agents busy at work, signalling when they need you. Each bot keeps its logo's shape and colours and follows your pointer with its eyes; Claude Code's is Clawd. See each bot's four states.

Small footprint. Light on your Mac.

A HUD runs all day, so staying light is part of the experience. Agent HUD keeps its CPU, memory and energy use down, and leaves your system to your agents and the work in front of you.

Measured on a real Mac, not claimed:

Measure Idle Busy Made up of
CPU, share of one core about 7% 7%–⁠28% Agent HUD 6–11%, each display's HUD 0.5–7%, graphics and networking up to 5%
Memory about 240 MB 240–⁠260 MB Agent HUD's own process, as Activity Monitor shows it. Each display's HUD adds 70–170 MB, graphics 10–50 MB and networking 7 MB
Energy over an 8-hour day about 0.1 Wh 0.1–⁠2.3 Wh Agent HUD more than half, the HUDs and graphics the rest

Idle: no agent working, one display. Busy: agents writing to their logs, one or two displays. Release builds on an Apple M1 Max, macOS 26, with about 8,000 session logs. How it is measured.

Built for every desktop

Agent HUD is built with Tauri 2: a Rust core, a React interface, and a thin native layer for each operating system.

  • One core. Reading your agents' records, counting tokens and working out quotas, prices and sessions happen once, in Rust, the same way on every system. The windows only display what the core computes.
  • Each system in its own files. Paths, processes, credentials, hooks and windows are written once per system, so work on Windows or Linux never touches shared code.
  • macOS today; Windows and Linux in preview. Both build from this source, and each release carries unsigned test packages for them. The Linux build has not been tried on a real desktop yet; what is still missing is under Coming soon.
  • iPhone and Apple Watch pair with a Mac or a Windows PC. Agent HUD Remote syncs through iCloud, so for now it is available on macOS only; with Windows, the phone connects on the same network.

Agent HUD Remote.

Your agents, wherever you are. Through your own iCloud, Agent HUD Remote keeps your iPhone and Apple Watch connected to the release on any network.

What you'll get

When your iPhone and the release share a network, iPhone and Apple Watch show your Mac's quotas, token statistics and sessions without a subscription. Remote adds:

  • Any network: your quotas, token statistics and sessions, away from home too.
  • Conversations: read Claude Code, Codex, Antigravity and DeepSeek Harness sessions, and message Claude Code and Codex.
  • Decisions: allow or deny a tool call, answer a question, approve a plan.
  • Your wrist: an idea you say becomes a new session on your Mac, and the reply comes back to your watch.
  • Alerts and Live Activities: alerts on iPhone and Apple Watch, and your running sessions on the Lock Screen and in the Dynamic Island.
iPhone Apple Watch
Quotas, balances and token statistics ✓ ✓
Sessions and their tokens ✓ ✓
Widgets on the Home Screen, Lock Screen and StandBy ✓ —
Several Macs ✓ ✓
Conversations ★ ★
Messages to Claude Code and Codex ★ ★
A note that becomes a new session ★ ★
Tool calls, questions and plans ★ ★
Alerts ★ ★
Running sessions as Live Activities ★ ★

✓ Free to use · ★ Needs a subscription

Check out in the App Store, inside the iPhone app: China mainland · International

Coming soon

  • Windows and Linux out of preview. Contributions are welcome.
  • More agents, beyond the 27 in the table below.
  • API keys billed by the token: their usage, spend and balance, next to your plans' quotas.
  • UI prompt templates. Ready-made prompts that reshape your HUD in one go.
  • A Gantt view of sub-agents. On a session's page, each sub-agent of a turn in its own lane on one time axis: when it started, how long it ran, what it used, and what ran side by side.
  • Agents shape every screen, not only the HUD, with image backgrounds and a customizable iPhone app.
  • More agents in the conversation, beyond Claude Code, Codex, Antigravity and DeepSeek Harness.
  • Files both ways. Your agents send files to your iPhone, and you send files from your iPhone to your agents.
  • Recordings straight to your agents. A recording from your Apple Watch or iPhone reaches your agent as it is.
  • More of the HUD is your call: what Simple mode's rings show (burn rate, time to reset or cost), and which agents may alert you.
  • More from each agent: Grok CLI's one click back, OpenClaw's live status and replies, Devin's and Goose's live status, and Goose's replies.
  • Requests and replies on the HUD in builds from source, once client hooks leave the preview.
  • HarmonyOS. An app for HarmonyOS phones is in review: on the same network as the release, your quotas, token statistics and sessions, with service cards on the home screen.

…and more.

Supported agents

Install and sign in to the agents you want to follow; Agent HUD finds them by itself.

Agent Tokens and sessions Quota Live status Approvals Questions Replies One click back Turns and calls Agent messages
Claude Code ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
Codex ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
DeepSeek Harness ✓ Balance ✓ ✓ ✓ ✓ Window ✓ ✓
Antigravity ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
Cursor ✓ ✓ ✓ ✓ ✓ Limited
Grok CLI ✓ ✓ ✓ ✓
Grok Bot ✓ ✓ ✓ ✓ ✓
GitHub Copilot CLI ✓ ✓ ✓ ✓ ✓
OpenCode ✓ ✓ ✓ ✓ ✓
Kilo Code ✓ ✓ ✓
MiMo Code ✓ ✓ ✓
Kimi ✓ ✓ ✓ ✓ Limited
Pi ✓ ✓ ✓ ✓ ✓ ✓
Oh My Pi ✓ ✓ ✓ ✓
OpenClaw ✓
Hermes Agent ✓ ✓ ✓ ✓
Factory Droid ✓ ✓ ✓ ✓ ✓
TRAE Sessions ✓ ✓ ✓ Limited
Qoder ✓ ✓ ✓ ✓ ✓ ✓ Limited
ZCode ✓ ✓ ✓ ✓ ✓ Limited
CodeBuddy ✓ ✓ ✓
WorkBuddy ✓ ✓ ✓ ✓ Limited
Qwen Code ✓ ✓ ✓
Devin ✓ ✓ ✓ ✓ ✓
Cline ✓ ✓
Roo Code ✓
Goose ✓

Limited messaging means a message goes in only when one of the session's turns ends: an idle session does not get it until you start another turn there.

Also: the GLM Coding Plan, OpenCode Go, MiniMax and Volcengine Ark plan quotas. Qoder includes Qoder CN and QoderWork, TRAE includes TRAE CN, and Devin includes the CLI and Devin Desktop. Codex CLI approvals need its hook trusted in Codex (the Codex app's need nothing), DeepSeek Harness approvals and questions need Agent HUD's plugin (Settings → Agent plugins), and Copilot's approvals and quota are each turned on in Settings. Qoder's, WorkBuddy's and TRAE's quotas show once you sign in to the vendor's site in Agent HUD's own window. TRAE reports no tokens, so only its sessions show, and it runs Agent HUD's hooks once you turn on its global hooks. Qoder records tokens, and so its turns, only for a model on your own key or with QODER_EXPOSE_TOKEN_USAGE set. A jump back to a terminal pane needs iTerm2, and Cursor opens a cloud agent's chat but only a local chat's folder. DeepSeek Harness's one click back brings its desktop app's window forward on macOS and Windows, not the session itself; on Windows the link already names the session, for when Harness supports it. An agent marked Limited under Agent messages takes a message only when one of its turns ends, not while idle, and cannot be started in a new session; only an interactive Qoder CLI session takes one while idle. OpenCode 2, Qoder CN and QoderWork take agent messages without sending them. Factory Droid's quota needs a Factory API key, and Devin's quotas come from its CLI's local cache. Cursor's tokens come from Cursor 3.27 or later. Grok Bot's bots run on Cursor's servers, which report each bot's daily tokens; an account merged into SuperGrok reports none, so only its sessions show. Grok Bot's weekly quota comes from the same servers when Cursor is signed in to the Bot's account, otherwise from the Bot's own cache, which it refreshes only while in use. An account merged into SuperGrok has one weekly limit for Cursor and Grok Bot, and Cursor's block shows it as Cursor does.

Keep your data safe.

  • Read on your computer: the records your agents already keep. Usage history stays in a local SQLite database.
  • Read from each vendor: quotas, balances and Grok Bot's daily tokens, with the agent's existing sign-in or configured key. No read sends a message to a model. Credentials are never stored, apart from a sign-in you make yourself on a vendor's site in Agent HUD's own window, which stays on this computer.
  • Written on your computer: hooks in your agents' own settings, to receive requests and replies, only while client hooks are on. Turning them off removes them.
  • Uploaded: nothing, by a build from this source.
  • Access your data from iPhone and Apple Watch, through the release:
    • On the same network, over HTTPS, with the Mac's certificate pinned when the phone pairs.
    • Anywhere, with Agent HUD Remote, through your own iCloud. Record contents are end-to-end encrypted with keys in your iCloud Keychain; only the text an alert shows stays outside that encryption, so iOS can display it.
    • Live Activities of running sessions, also with Remote, pass through Agent HUD's relay to Apple's push service. The relay keeps only what forwarding needs, each Mac's latest sessions and each phone's card, with no account identity, and deletes an account it hasn't heard from in 30 days.

Community

Questions, ideas and bug reports are welcome on Discord, in the WeChat group, or in issues.

WeChat group: scan the code with WeChat

A group code lasts a week, and I replace it on agenthud.app each week; this picture follows it. If it has expired, find me on Discord.

Build from source

Requirements: Node.js 24+, pnpm, a current Rust toolchain (1.90+), and Tauri's prerequisites for your system.

git clone https://github.com/jazzenchen/agent-hud-open.git
cd agent-hud-open
pnpm install
pnpm run check
pnpm run build:macos -- --bundle

Use build:windows or build:linux on those systems; each command builds only for the system it runs on. Tauri writes the app under target/release/bundle. To try the app without any agents, start it with --demo, which opens it on sample data.

A patched Tao. On macOS the HUD has to be a panel that never takes focus from the app you are using, and upstream Tao cannot create one yet (tao#414). Until it can, the workspace takes tao from a public fork with one small commit on top of 0.37.1, pinned by revision in Cargo.toml; Cargo fetches it from GitHub on the first build. What the commit changes, and how to move it to a new Tao release: Tao patch.

Preview. This build is checked against the native Swift app, agent-hud-native-open (formerly agent-hud-open), behavior by behavior, in the parity ledger. It does not install client hooks yet, so approvals, questions and most replies stay in each agent's own prompt.

Development
Command Purpose
pnpm run check Build the interface and MCP server, check the platform boundary and the parity ledger, type-check Rust
cargo test --workspace Run the Rust tests
pnpm run dev Run the interface in a browser at http://127.0.0.1:1420 on live data (run pnpm run engine:build first)
pnpm run build:macos / build:windows / build:linux Build the app for the system you are on
pnpm --silent run mcp Run the MCP server over stdio

The workspace layout, the MCP client configuration and the other commands are in development. Code goes where Contributing says: core computes, the interface displays, and each system's code stays in its own files.

Documentation

Development · Platform architecture · Platform builds · Accounts · Client hooks · Session data · Local network · Performance · Parity ledger · Tao patch · Third-party sources · Contributing

License

Apache-2.0. Interface components, provider artwork and Clawd's sprites keep their own notices; see third-party sources.

Star History

排名相邻

AI Rank iPhone 版

随时追踪 AI 模型榜单

在 App Store 下载