See what your Claude Code agents are actually doing, while they do it.
When a session delegates, the work disappears. You see "Agent launched successfully" and then nothing until it returns. Agent Max puts the whole tree on screen while it runs, so you can watch every agent work, see what it is holding against its context window, and know what it has spent.
Install
From VS Code: open the Extensions view, search Agent Max, click Install.
From the command line:
code --install-extension Davidarias01.agent-max
Opening it
Install it, then click the button at the bottom left of VS Code. That is the whole thing. The item says Agents, it sits in the status bar next to the errors and warnings counts, and clicking it opens the exhibit beside your code.
If you do not see the button, restart VS Code. A freshly installed extension does not always activate in the window you installed it from. One reload and the button is there for good.
During a run the button turns into a live count, 2/8 agents, so you can tell
work is happening without opening anything. agentMax.statusBar removes it if
you want the space back.
Other ways to open it
The status bar is the quickest, but nothing depends on it. Any of these gets you to the same place, and the watcher starts on first open whichever one you use.
From the bottom panel. There is an Agents tab next to Terminal, Problems and Output from the moment you install. Nothing to run, it is already there. This is the one to use if you like keeping it docked under your code.
From the panel to the right. In the Agents panel's title bar there are two icons. The split one moves the exhibit into the column beside your editor, where it behaves like any other tab and can be split, dragged or maximised. The other restarts the watcher.
From the command palette. Ctrl+Shift+P, or Cmd+Shift+P on a Mac, then:
| Type this | What you get |
|---|---|
Agent Max: Open Beside Editor |
The exhibit as an editor tab to the right |
Agent Max: Show in Bottom Panel |
Focus the panel view instead |
Agent Max: Open in Browser |
The same page in your browser |
In a browser, on a second screen. Agent Max: Open in Browser serves the identical page outside VS Code, which is the one to use if you want the run on a monitor of its own while you keep coding on the main one.
With your own keyboard shortcut. No default shortcut ships with the
extension, deliberately, because any combo worth binding is already bound to
something in somebody's setup. To pick your own:
Ctrl+K Ctrl+S, search Agent Max,
then click the + beside the command you want.
Moving it after the fact. Right-click the panel and Move Panel Right if you would rather keep it as a panel but on the side. VS Code remembers where you put it.
Requires Python 3.9 or later on your PATH. The watcher is a single standard-library script with no pip dependencies, bundled with the extension. It is found automatically (
py -3on Windows, thenpython3, thenpython); if yours lives somewhere unusual, setagentMax.pythonPath.
Features
Workflow runs, with their phases
A workflow is not an ad-hoc agent tree. It declares its phases up front, its agents get labels from the script, and it keeps a status after every agent is gone. So it gets its own tab.
Each run shows a timeline built from real timestamps, grouped by the phase that spawned each agent, which is how you see whether a stage pipelined or waited. Overlapping bars mean concurrency; one long bar starting late means a barrier. Click any agent to open the same card the tree opens.
A run still in flight has written no record yet, so its name, phases and timing are rebuilt from the transcripts. Its bars run to now and grow while you watch, working agents pulse, and each row names the tool it is running this second.
Runs are scoped separately from the agent tree, because the tree is about this session and a run you want to read usually is not.
The tree, to any depth
Agents that spawn their own agents nest under them. A run three levels deep looks three levels deep, and each card carries context against the window, turns, tool calls and output tokens.
What each agent is doing, right now
Agents appear the moment they are created, with the brief they were given.
Last said is the most recent line from that agent. Last action is the last tool call it made, with a timestamp. Between them you can tell a stuck agent from a slow one without opening anything.
Session totals, kept apart
| Context in play | size of the current prompt. A snapshot, not a running total |
| Written, all turns | output tokens the session actually produced |
| Biggest single call | the largest one, useful for spotting a context blow-up |
| Billed, all turns | the cumulative cost, high because every turn re-sends its context |
Filter the agent list by session, depth, model, or down to agents that did real work, which is the fastest way to hide the ones that spawned and returned immediately.
Dialogue, log and trace
Open any agent and read it back. Everything it said with timestamps, every tool call in order, and the raw trace underneath. You can follow a number in the output back to the call that produced it.
Commands
| Command | Description |
|---|---|
Agent Max: Open Beside Editor |
Open as an editor tab to the right |
Agent Max: Show in Bottom Panel |
Focus the panel view |
Agent Max: Restart Watcher |
Stop and restart the local process |
Agent Max: Open in Browser |
Open the same page outside VS Code |
Agent Max: Show Logs |
Output channel, when something is wrong |
Settings
| Setting | Default | Description |
|---|---|---|
agentMax.pythonPath |
"" |
Python 3.9+ interpreter. Empty means auto-detect |
agentMax.port |
8780 |
Localhost port the watcher listens on |
agentMax.limit |
16 |
Maximum transcripts watched at once |
agentMax.recentMinutes |
90 |
Hide agents idle longer than this. 0 shows all |
agentMax.statusBar |
true |
Show the status bar item, with a live count during a run |
agentMax.autoStart |
true |
Start the watcher when the view opens |
agentMax.watcherDir |
"" |
Advanced. Point at your own copy of the watcher |
agentMax.limit is worth knowing about. Agents past the cap are hidden without
warning, so if a run fans out wider than you expected, raise it.
Privacy
Everything is local. The extension makes no network requests.
The watcher reads Claude Code's own transcript files under ~/.claude/projects/,
which is where your sessions already live, and serves a page on 127.0.0.1.
Nothing is uploaded, no API keys are read, no telemetry is collected, and no
data leaves the machine. The port is bound to the loopback address only, so it
is not reachable from your network.
The watcher is a single readable Python file:
watcher/watch_agents.py.
Known behaviour
- If the port is already serving, the extension reuses it and does not start a second watcher. Two processes on one port means the second binds nothing and exits quietly, which looks identical to the extension being broken.
- Closing the last VS Code window stops the watcher it started.
Contributing
Issues and pull requests welcome at github.com/davidalmeida90/agent-max.
To run from source: clone the repo, open it in VS Code, press F5 to
launch an Extension Development Host. To build a .vsix, run npm run package.
Not affiliated with Anthropic
An independent tool that reads local Claude Code transcripts. Not built, endorsed or supported by Anthropic.
License
MIT © David Arias