live·Claude Code · Codex · OpenCode·0 network calls

See what your coding agents are doing. Right now.

Agent Deck is a free extension for VS Code, Cursor, Devin (Windsurf) and Kiro that draws a live tree of every agent and every tool call, as they work. Claude Code, Codex and OpenCode.

Install for VS Code → Install for Cursor, Devin or Kiro → live on the VS Code Marketplace and Open VSX

Works in

VS Code
Cursor
Devin (Windsurf)
Kiro

Agent Deck drawing a live tree as one session spawns 4 lead agents, each spawning 2 workers — live capture · real session1 session → 4 leads → 8 workers

liveLive, not after the fact.

It updates while your agents work, not when the session ends.

read-onlyIt watches. It never touches.

Never starts, steers or changes your agents or their settings.

0 network callsNothing leaves your machine.

No telemetry, no cloud, no account. Free and open source (MIT).

~/what-is-it

Live observability for AI coding agents.

In plain words: a live tree of your agents and their tool calls, right inside your editor.

editor extensionFree and open source (MIT). For VS Code, Cursor, Devin (Windsurf) and Kiro.
works withClaude Code, Codex and OpenCode. You run them as usual; Agent Deck only watches, and shows their sessions side by side in one panel.
session filesReads the files your agents already write to your disk as they work.
hooksOptional. Claude Code and Codex hooks ping it on 127.0.0.1 when an agent starts or a tool finishes. You paste them in yourself.

~/problem

Your agents work in parallel. Your terminal shows one line at a time.

Give Claude Code or Codex a big task and it splits the work across several agents running in parallel. In the terminal you can barely see who's doing what, who's stuck, or why it's taking 10 minutes. The logs exist, but as raw files nobody reads while the work is running.

"I built Agent Deck because I couldn't understand what was running on my own machine."
Nadav Vitlam, creator

~/where-it-sits

On the side, reading. Never in the way.

Agent Deck isn't in your terminal and isn't between your agent and the API. It reads the files your agent already writes, listens to hooks, and draws what it reads.

solid blue: reads · dashed blue: optional hooks · red ✕: never

readsThe session files your agent already writes to disk, as they grow. Read-only.
listensOptional hook signals you paste into your agent's settings, sent to 127.0.0.1 only.
neverSits in the API path, wraps or steers your agent, writes to its files, or sends anything out.

~/what-you-see

One panel. Everything your agents are doing.

stalledStuck agents, spotted.A call silent for too long turns amber. When an agent is waiting for your answer, it says so.
cost.sourceEvery cost names its source.The agent's own figure, or your own prices. No guessed price tables.

~/who-its-for

For developers who hand agents long tasks.

Solo developers and small teams who give Claude Code, Codex or OpenCode long tasks with subagents. It isn't a management dashboard and it isn't enterprise monitoring.

seeing firstThe point is seeing what's happening: who's working, who's stuck, where the time goes.
tokens secondTokens and cost are shown too. Even on a subscription they help you stay inside your usage limits.

~/why-agent-deck

Other tools tell you what happened. Agent Deck shows you what's happening.

after the factUsage reports and transcript viewers work once the session is over. They show you the bill, not what's running now.
inside the agentYour agent's own window shows its conversation, not the whole tree of agents working under it.
agent deckEvery agent and every tool call, live, in one tree. Read-only, and no network.

~/reviews★★★★★ on the Marketplace

“Agent Deck fixes the endless terminal scroll.”
Nir Tituani

“It feels like getting a bird's-eye view of the entire environment.”
Alex k

read all reviews →

Want to know what to change next time?

Agent Deck Insights is optional and paid. Your own Claude Code or Codex reads the stats Agent Deck collected and tells you what to change. Agent Deck stays free.

See Insights →

example finding → Split long phases into shorter sessions, handing over through a written file.

~/installabout 2 minutes

Up and running in three steps.

VS Code: install from the VS Code Marketplace, or run

code --install-extension nvitlam.agent-deck

Cursor, Devin (Windsurf) or Kiro: install from Open VSX

Click the Agent Deck icon in your editor's sidebar. Your sessions appear.

Paste one hook blockoptional into your Claude Code or Codex settings for second-by-second updates. You paste it yourself, because Agent Deck never edits your settings. Guide →

~/faq

What does "observability" mean here?

Not OpenTelemetry. It's closest to live traces: a tree of agents and tool calls, with the time and token count on each one. There's no log export and no Grafana metrics. It's a local, live viewer, not monitoring infrastructure.

Is this about saving tokens?

Mostly not. The main point is seeing what's happening. Tokens and cost are shown, and they help with usage limits, but they come second.

Can it control my agents?

No. It only watches.

Does it send my data anywhere?

No. It makes no network calls at all.

Why do I paste the hook myself?

Because read-only includes your settings. Agent Deck never writes to them.

What are the limits?

The session file format is undocumented and changes between versions. When it changes, Agent Deck refuses to draw rather than guess, so an agent update can break it until I fix it. It's a one-person project, and it's early.

Which versions does it support?

See the compatibility details in the README.