Supported agents

Agent HUD follows 18 coding agents, plus the GLM Coding Plan quota and permission requests from the Qoder builds. Install and sign in to the agents you want to follow; Agent HUD finds them on your Mac by itself. What it can show for each one depends on what that agent records and reports.

What each agent supportsCopy Anchor Link

Also: the GLM Coding Plan quota, and approvals from Qoder, Qoder CN and QoderWork.

• Tokens and sessions: each session with its tokens, models and estimated cost. See Tokens and cost.
• Quota: the windows or balance the vendor reports. See Quotas and balances.
• Live status: the HUD shows the agent working, waiting for you or done.
• Approvals and Questions: answered on the HUD. See Requests.
• Replies: a New reply reminder on the HUD when a turn ends.
• One click back: opens the session in the agent's app or terminal pane.
• Turns and calls: a session's page breaks it down by turn and by call.

What needs setting upCopy Anchor Link

For each agent, Settings → Agents also has a Live status switch and a Completion reminders switch. With live status off, the agent's sessions show no state or last message and bring no reminder; its tokens and history keep counting.

Notes by agentCopy Anchor Link

Claude CodeCopy Anchor Link

Agent HUD reads the transcripts Claude Code keeps in its projects folder. For the plan limits, it asks the installed Claude Code for the signed-in account's usage, with hooks off; no message goes to a model. The account shows with its email and plan; a Max plan shows its 5× or 20× tier. An API key or a third-party platform has no plan limits, and the quota block says so. One click back opens the session in the Claude desktop app, or in its iTerm pane when it runs outside tmux. Questions Claude asks in its desktop app are answered through Agent HUD's plugin for Claude Code, which you install from Settings → Agent plugins.

CodexCopy Anchor Link

Both the Codex app and the CLI. Agent HUD reads Codex's session logs, its archived sessions and its session index. It reads the limits from the installed Codex through Codex's own app server, without starting a thread or a turn. The app and the CLI on one sign-in show one set of windows. CLI approvals reach the HUD only for turns that ask for your review; what Codex reviews by itself stays in Codex. One click back opens the thread in the Codex app, or the iTerm pane a CLI session runs in outside tmux. To follow more Codex accounts, see Quotas and balances.

DeepSeek HarnessCopy Anchor Link

DeepSeek bills a balance, so Agent HUD shows the balance, the spend per day and the day it runs out, instead of a quota window. It reads the balance with the key Harness is configured with, and sends that key only to DeepSeek's own balance service. With a custom API address there is no official balance to read: Agent HUD sends no request and says so. Questions and approvals reach the HUD through Agent HUD's plugin for Harness; whichever you answer first, on the HUD or in Harness, counts.

AntigravityCopy Anchor Link

The quota comes from the Antigravity app running on this Mac and from connected Antigravity CLIs. Agent HUD keeps any status line you already use in the CLI. Each email is its own account; a reading older than five minutes is marked stale until a new one arrives. Agent HUD answers only the approvals and questions the running app is already asking. It does not show Antigravity working during a turn. One click back opens the conversation in the running app and never starts the app itself.

CursorCopy Anchor Link

Sessions come from Cursor's own database on this Mac, and their tokens from Cursor's completion hook, which reports each generation's tokens from Cursor 3.27 on. The quota comes from your Cursor account's usage summary: each part of your plan the account reports, with the reset Cursor gives. Approvals and questions stay in Cursor.

Grok CLI and Grok BotCopy Anchor Link

Two separate sources. Grok CLI: sessions and tokens from ~/.grok, or the folder GROK_HOME names; its quota is not read, and no Grok credential is read. Grok Bot: the quota comes from the cache the Grok Bot app keeps on your Mac, and its sign-in files are never opened. Its bots run on Cursor's servers, so each bot's daily tokens are read from there with Cursor's sign-in on this Mac. An account merged into SuperGrok reports no tokens, so only its sessions show.

GitHub Copilot CLICopy Anchor Link

Tokens and sessions come from Copilot CLI's own folder. The quota is read only after you turn on Read GitHub Copilot quota, using the GitHub CLI's existing sign-in. Agent HUD launches neither gh nor Copilot and never refreshes the token. A quota that is unlimited on your plan gets no row.

OpenCode, Kimi and PiCopy Anchor Link

• OpenCode: the OpenCode Go plan's limits, read with the key OpenCode already holds. A Kimi Code or GLM Coding Plan key held in OpenCode shows that plan's limits too. Replies come from Agent HUD's OpenCode plugin. One click back works when OpenCode runs in iTerm outside tmux.
• Kimi: the limits of your Kimi Code plan, read with Kimi Code's existing sign-in or the key in KIMI_CODE_API_KEY.
• Pi: has no quota of its own; you see the plan that pays for its requests. A ChatGPT sign-in in Pi appears as Codex's quota, counted once with Codex. A GLM Coding Plan key appears as the GLM quota. Live status and replies come from a small extension Agent HUD adds to Pi. One click back works for iTerm2 sessions outside tmux.

OpenClaw and Hermes AgentCopy Anchor Link

Tokens and sessions only, read from the files each one keeps: OpenClaw's state folder (~/.openclaw or OPENCLAW_STATE_DIR) and Hermes's state database (~/.hermes or HERMES_HOME), every profile included. Agent HUD reads no messages from Hermes, no credentials from either, and adds nothing to their settings. When OpenClaw runs a turn through Claude Code, the turn counts once, under Claude Code.

ZCode, CodeBuddy, WorkBuddy, Qwen Code and Devin CLICopy Anchor Link

Their tool approvals reach the HUD with Deny and Allow once; questions and plan approvals stay in the agent, except ZCode's questions, which reach the HUD too.

• ZCode runs hooks only while its own hooks setting is on. When the setting is missing, Agent HUD turns it on with its hook; if you turned it off, Agent HUD leaves it off.
• CodeBuddy and WorkBuddy record their work the same way but count as two agents, so their sessions never mix. When WorkBuddy records that a model reached its plan's limit, that limit shows as used up until the reset time WorkBuddy gave.
• Qwen Code: a session started from an earlier one copies its history; Agent HUD counts that history once.
• Devin CLI: the daily and weekly limits come from the account status the CLI caches on your Mac. They are as fresh as the CLI's own last check; Agent HUD sends no request of its own.

On iPhone and Apple WatchCopy Anchor Link

Every agent's tokens, sessions and quota reach iPhone and Apple Watch: free when your iPhone is on the same network as your Mac, and on any network with Agent HUD Remote. Remote also brings the approvals, questions and plans the HUD receives. See iPhone and Agent HUD Remote.