Agent Standup Tracks Claude Code Without Extra Tokens
The unofficial dashboard visualizes Anthropic’s Claude Code session files locally, exposing agent activity without extra model calls unless optional AI features are enabled.
Listen
AI narration
10:32
0:00 / 10:32
AI SummaryGenerated from this article
Agent Standup, an unofficial open-source dashboard, visualizes Claude Code session files locally without requiring additional model calls. The tool displays agent activity, status, commands, and edits in a live animated interface, helping developers identify which agents are working, waiting, or need attention.
The project offers multiple installation options including VS Code extensions, terminal access, and experimental desktop builds for macOS. It tracks context usage and token consumption from session records, provides activity indicators with configurable thresholds, and includes replay features with timeline highlights. Optional AI features for French translation and jokes require the developer's own Claude credentials.
Claude Code already writes session records containing agent activity. Agent Standup reads those files and turns them into a live, animated team meeting, complete with agent status, commands, output, and edits. It does not need another model to summarize the work.
The meeting presentation is playful, but its practical use is straightforward: developers can see which agents are working, waiting on another agent, requesting attention, stalled, or finished. They get an overview alongside the ability to inspect individual agents, instead of watching a single stream of activity.
The project describes itself as free and open-source. It is explicitly unofficial, with no affiliation with or endorsement from Anthropic. Its “zero token cost” claim applies to ordinary visualization; French translation and AI-generated jokes can use the developer’s own Claude login.
This walkthrough follows the published documentation and does not include hands-on testing. That limits what it can establish about a tool that reads local development-session data and offers an unsigned desktop build.
Existing Session Files Become a Live Agent View
The Agent Standup project page describes one small local server with several interfaces: an editor panel, a browser-accessible terminal launcher, and a standalone desktop window.
Because the dashboard uses session data Claude Code already produces, it can observe existing work without needing prompts, context, or model calls to explain what happened.
In the live meeting view, the lead agent gives assignments and sub-agents appear as participants. Their shared screens show recorded work: terminal commands and output, search results, file contents, and edit diffs. Clicking an agent exposes its prompt, tool usage, reports, and recent activity.
A hierarchy view shows who delegated work to whom, while a Sessions board spans projects. These views give the animation a practical purpose: the tree explains delegation, and the status list helps identify where attention is needed.
The meeting remains a visualization of recorded activity. A character speaking, leaving, or celebrating is not independent confirmation that the underlying task succeeded.
Choose the Installation Route That Fits Your Workflow
The published installation options cover several environments, though some use the same package rather than a native integration.
Work with Zeniteq
Let’s work together
We’re open to thoughtful collaborations with teams building in AI. Explore the ways we can work together.
VS Code: Search for Agent Standup in the Extensions view or follow the project page’s Marketplace link. The extension provides a panel and sidebar. Its documented status-bar integration shows workspace counts, including agents working or needing attention, and can open the most urgent session.
Cursor, Windsurf, and VSCodium: Download the project’s .vsix package and use Extensions → … → Install from VSIX…. Each editor uses the same extension installed from a file, not a separately built product.
A terminal and browser: The README specifies Node.js 18 or later and provides this launcher:
npx @double_geste/agent-standup --open
This avoids installing an editor extension, but still executes a third-party package. The convenience of npx is not a security boundary.
To preview the interface without using your own session data, select Settings → Source → Demo. The built-in demo depicts a fictional team fixing a slow login page. It can help you learn the interface, though it provides no evidence of how reliably the dashboard handles your workload.
For real activity, the project says Claude Code must have been used on that machine at least once. The documented live view follows the newest Claude Code session; the Sessions board provides a broader project overview.
Desktop: The project offers macOS builds for Apple Silicon and Intel. It labels Windows and Linux desktop builds experimental and warns that the macOS app is not signed by Apple.
Claude Code status line or tmux: For a compact indicator, the documented commands are:
npm i -g @double_geste/agent-standup
agent-standup statusline
The project identifies agent-standup statusline with Claude Code and agent-standup status with tmux. Consult the README for integration details. Running either command should not be assumed to complete the host application’s configuration.
JetBrains users have a separate limitation: the project page says there is no native JetBrains plugin yet.
Watch Attention Signals Before Watching the Animation
The strongest workflow feature is the distinction between activity and attention. A dashboard showing ten busy agents is less helpful than one showing that a single agent needs the developer to respond.
Agent Standup advertises an urgency-sorted list and five broad states: working, waiting on another agent, needing you, stalled, and done. Those states can help a developer decide which session to inspect first.
The README’s activity indicators provide more detail. A tile pulses amber when a command has run for over a minute. A working agent that has gone quiet for three minutes gets a sleep indicator. Failed tool calls contribute to a red failure counter.
These thresholds flag activity worth checking without establishing its cause. A long-running command might be a legitimate test suite, and silence might reflect ongoing work rather than a deadlock. Before interrupting an agent, inspect the underlying command and output.
The shared-screen view supports that inspection. It shows what an agent is reading, editing, searching, or executing, along with tool status and duration. That can distinguish an agent waiting on a command from one repeatedly encountering errors.
There are display limits: the documentation says an agent’s recent-activity view contains its last 30 calls, with outputs trimmed to roughly 1–2 KB each. Internal-call results, including sub-agent launches and messages, are not shown there.
The dashboard therefore serves as a quick inspection surface, not a complete forensic record. If the visible output does not explain a failure, return to the underlying session or command output.
Context Gauges and Token Counters Answer Different Questions
Agent Standup’s context gauge and token counter may appear together, but they measure different things.
The context gauge reports how full the lead agent’s context is and provides a warning before auto-compaction. It lets a developer notice when a long session is approaching a context-management event, without establishing whether the agent will preserve every important detail through it.
The token counter tracks usage recorded in Claude Code’s session files. According to the documented counting method, it includes input, output, and cache writes, counted once per message, with cache reads shown separately in a tooltip. The project explicitly says this is not an estimate of the bill. A larger counter can identify a relatively token-heavy agent, but it does not supply an authoritative dollar cost.
For completed work, replay adds a timeline, highlights such as the first error and longest wait, and a recap covering agents, tool calls, commands, and changed files. Possible-loop highlights can direct attention to repetitive activity without requiring a linear reread of the whole session.
These features help with investigation. They cannot independently determine whether an edit was correct or whether a reported fix passed the checks that matter to the project.
Local Operation Still Requires a Trust Decision
The project says ordinary visualization runs locally and does not send code elsewhere. Its README also states that it does not call Claude unless French translation or jokes are enabled.
That defines the scope of the “zero-token” promise: the dashboard adds no model calls in its ordinary mode, while Claude Code’s underlying work still consumes tokens. Enabling the optional model-backed features changes that arrangement.
Local processing is a meaningful privacy property, though it is no substitute for a security audit. The tool intentionally reads session records and displays prompts, commands, outputs, file contents, and diffs. Depending on the work, those records can contain sensitive information. Installing it means trusting third-party software with that material.
The unsigned macOS build warning deserves similar care. Unsigned does not establish that an application is malicious. It does give you a reason to evaluate the download deliberately instead of dismissing an operating-system warning as a nuisance. Terminal and editor installations also require trust; avoiding the desktop build does not remove the need to assess the package.
Exports introduce another exposure path. The README documents a “Hide code & file names” option for videos, but says it is off by default. The option does not mask speech bubbles, subtitles, or voices. If an agent says a filename or quotes code aloud, visual masking will not remove that disclosure.
Review an exported clip before sharing it. A colorful agent meeting can still carry private development data.
Agent Standup’s most defensible value is reducing the effort of finding the next thing to inspect. Its status list, delegation tree, command view, and replay can make parallel Claude Code work easier to follow without assigning another model to narrate it. For developers willing to trust a community tool with local session records, that is a concrete benefit. Visible progress, however, is not verified correctness.
