Install

1. Search Claude Agent Map & Usage in the Extensions tab, or run code --install-extension pandaprog.claude-agents.
2. Click the Claude Agents icon in the activity bar.
3. That's it - every running Claude Code session shows up.

Nothing to configure. The extension reads ~/.claude on the same machine, read-only, and sends nothing anywhere.

What you see

• Live activity per session - 💭 thinking, ✎ editing, ⏵ running, 🤖 delegating.
• Every agent of the session, finished ones included, with the same duration and tokens as Claude Code.
• Orange when a session asks you something, with a badge on the activity bar icon.
• Context bar and model, prompt cache ⏱, effort level.
• Background commands with their live localhost links.
• Failed agent? Hover it to read why.
• Superpowers plans - each task of a subagent-driven plan as a square, with its agents' time and tokens.
• Session statistics - ⤓ exports every agent of a session to CSV, with its tokens and what it would have cost on the Claude API.
• Cost on hover - hover a task or agent line for its agents and API cost, 🤖×2 · 1.84$ · 42 min · 610k; click any of these lines to show them everywhere, click again to go back.
• Idle sessions step aside - after 10 minutes without activity, a session folds down to its live localhost links, or leaves the view if it has none. The window opened on that session's folder keeps it in full.
• Light, even with many windows open - the scan runs off VS Code's main thread. A window without the focus refreshes every 10 s and pauses its animations, and a hidden view only keeps the badge up to date.

Session statistics

Click ⤓ at the end of a session's title line for a CSV of the whole session: the main agent, then every sub-agent and workflow agent, and a total line.

• Tokens by kind - input, cache writes (5 min and 1 h), cache reads, output - summed from the usage Claude Code records for each API call.
• API cost, kind by kind - every call priced at its own model's public API rate (September 2026 price list), so a session that switched models is priced right.
• Nothing guessed - a model missing from the price list, or fast mode, leaves the cost cells empty and the export tells you how many agents are concerned.
• Opens straight in Excel - semicolons and decimal commas with VS Code in French, commas and points otherwise.

Superpowers plans

A session running a superpowers subagent-driven plan gets a 📋 Plan block: one square per task, the done/total count, and when the plan was last written. Tasks not started yet show as empty squares, so you see from the start how far the plan goes. A task added along the way counts in the total, titled from its brief.

• Task details - click 📋 Plan or the squares for the list: ✅ done · 🔵 in progress · 🟡 in review · ⬜ to do, each task with its title and 42 min · 610k - the time its agents worked on it, from the first start to the last finish, and their tokens. Hover it for 🤖×2 · 1.84$ in front: how many agents, and what they would have cost on the Claude API.
• Hover a task for its state and what its workspace holds: brief, report, review, fix round.
• Browse plans - ‹ › step through the folder's earlier plans, newest first; › all the way brings you back to the current one.
• Finished plans step aside - once the final review is there (✓) and you talk to Claude about something else, the block folds into a 📋 N plans line that keeps them one click away.

Plan usage

Your claude.ai limits at the bottom of the view, fed by Claude Usage, a small tray app that reads them for you - no API key.

1. Install Claude Usage and log in to claude.ai.
2. Copy the Output file path from its settings.
3. Paste it into claudeAgents.usageFile. The gauges appear.

Your limits also sit at the far right of the status bar, with a live countdown to each reset: Session 2% 4:09:21 · Weekly 18% 3d 04h. A limit turns yellow at 80% and red at 95%. The small ⌄ next to it switches to rings, moves it to the left, hides it, or toggles the card in the view.

With several claude.ai accounts in Claude Usage, the card and the status bar tooltip name the one the gauges belong to (its email for a personal plan). The name only shows once Claude Usage has written that account's reading, never right after a switch.

Settings

What's new in 0.9.4

• Which account - the usage card (USAGE · you@example.com · updated 7 min ago) and the status bar tooltip name the Claude Usage account the gauges belong to.
• Plans - tasks not started yet show as small empty squares, and a task added to a plan along the way counts in its total (no more 15/14), titled from its brief.
• A spinning hourglass while the first scan runs, instead of "No Claude session" before anything was read.
• Plain hyphens everywhere the extension writes: no more long dashes in labels, tooltips or the CSV.
• Much lighter on big sessions - a session with hundreds of agents used to freeze the extension host for seconds at start-up and cost a slice of CPU every 2 s.
  • The scan now runs in a worker thread, which only starts once the view is opened.
  • It remembers what it read across restarts, in a cache in VS Code's global storage checked against each file's modification time.
  • It checks a settled agent every 30 s instead of every 2 s, opens each transcript once, and no longer searches every sub-folder of your project for plans on each scan.
  • A resumed session (claude --resume) is no longer read again from scratch.
• Many windows open - each window now uses about half the CPU and a quarter less memory:
  • A window without the focus refreshes its cards every 10 s instead of 2 s. Durations keep ticking, and the cards catch up as soon as you click back in.
  • A hidden view only reads the sessions themselves, to keep the badge right: no agent transcript is touched until you open it.
  • Pulsing dots and squares pause in a window without the focus, and when your system asks to reduce motion.
  • The view redraws once per second instead of four times, only gets a new state when something changed, and no longer receives the sessions it hides. Nothing is redrawn in a minimized window.
  • The retention ring of a finished agent moves once per second instead of animating continuously.
  • The shared cache is half its former size, and only one window at a time rewrites it, once an agent has settled.
• Faster CSV export - it no longer loads every transcript at once: exporting a session of hundreds of agents takes a fraction of the time and memory.

What's new in 0.9.3

• Session statistics - ⤓ at the end of a session's title line exports every agent of the session to CSV: tokens by kind (input, cache writes 5 min and 1 h, cache reads, output) and what each would have cost on the Claude API, kind by kind, with a total line.
• Cost on hover - task and agent lines show time and tokens; hover one to add its agents and API cost (an agent counts the sub-agents it launched), or click to show them on every line. The cost stays blank until every transcript involved has been read and priced.

What's new in 0.9.2

• Marketplace rating in the top banner - the real average as stars, the number of reviews on their left.
• The stars now light up one by one, go dark, then show the real rating.

What's new in 0.9.1

• Superpowers plans - a session running a superpowers subagent-driven plan shows its tasks as squares: done, running, in review, to do. Click 📋 Plan or the squares for the list.
• Time and tokens per task - each task adds up the agents that worked on it, from the first start to the last finish.
• Finished plans step aside once you talk to Claude about something else; ‹ › browse the folder's earlier plans.
• Agent rows: the current activity sits right after the title, and the agent type moved to the tooltip.

What's new in 0.9.0

• Agent map - click N agents for every agent of the session, with exact duration and tokens.
• Plan usage - gauges at the bottom of the view, with the new Claude Usage app.
• Plan usage in the status bar - session and weekly limits with a live countdown to the reset, or as rings.
• Per-session effort, prompt cache ⏱, and why an agent failed on hover.
• Settings gear in the view title, next to the project filter.
• Sessions of the open workspace stay visible.

Privacy

Read-only on ~/.claude, never touches .credentials.json; to name the usage account it also reads Claude Usage's own settings and account list (the output file path, the active account's name), nothing else. The only files it writes are the statistics CSV you ask for, where you choose, and its own scan cache in VS Code's global storage for the extension (what it last read from your transcripts, so the next start does not read them again; a transcript is forgotten 48 hours after it was last seen). Two network calls only: a TCP check on 127.0.0.1 to tell a live dev server from a dead one, and - while the view is open, at most every 6 hours - a read of the Marketplace's public reviews to show the extension's rating. Nothing is sent.