Local dashboard for visualizing usage across eight AI coding agents — sessions, token costs, cache performance, tool calls, and daily breakdowns. Filter by agent or view everything in one place.

Inspired by foyzulkarim/claude-lens.

Multi-Agent Usage Dashboard

OpenCode and Cursor read SQLite via the optional better-sqlite3 dependency. If the prebuild fails on your platform the dashboard still works for the JSONL/JSON-based agents (Claude, Codex, Gemini, Kimi, Antigravity).

• Agent selector — filter the entire dashboard by agent or aggregate across all
• Sticky topbar — animated brand mark, live KPI pills (today cost / today messages / total cost with day-over-day trend), refresh button, section-nav pills with scroll-spy
• Today vs All-Time stats — sessions, messages, tool calls, estimated cost, cache hit rate
• Charts & Metrics — pure inline-SVG (no deps), responsive via ResizeObserver:
  • Daily Cost line chart
  • Token Volume stacked bars (input / output / cache read / cache write)
  • Activity by Hour and Weekday bars
  • Agent / Model / Tool distribution donuts
• Cache performance — hit rate, savings vs no-cache baseline (uses the active agent's pricing)
• Daily cost & cache table — per-day token breakdown with estimated spend
• Tool call analytics — which tools each agent used most, click any tool for the call list with export
• Prompt history — searchable, project-tabbed, copy-to-clipboard per row
• Project activity — searchable, agent badges, last-active relative time
• Collapsible lists — projects / tools / history / daily-costs start collapsed with a gradient fade and a "Show all" pill (initial caps: 8 / 8 / 25 / 7)
• Live OpenRouter pricing — refreshes per-model rates in the background; env vars override
• Per-agent pricing — independent rate config per provider via .env
• Polish — skeleton shimmer loaders, friendly empty states, sticky table headers, zebra rows, tabular-num alignment, prefers-reduced-motion support

The dashboard is plain JSON over HTTP — useful for scripting or integrations.

All endpoints accept an optional agent query param to scope by adapter name.

• Node.js 18+
• One or more of the supported agents installed locally
• Works on macOS, Linux, and Windows. Default data directories are auto-detected per OS.

The dashboard runs on macOS, Linux, and Windows. Path resolution per agent:

• Verified: macOS — actively tested against real installs.
• By convention: Linux/Windows — adapters use os.homedir() + the agent's documented dotfile, or the standard Electron app.getPath('userData') location for IDE-style agents (Cursor, Copilot via VS Code, Antigravity). These match each agent's published behavior; report mismatches via issue with the actual path your install uses.
• Always overridable: every adapter accepts an env var to point at any path you want.

Override any of these with CLAUDE_DIR, CODEX_DIR, GEMINI_DIR, OPENCODE_DIR, KIMI_DIR, CURSOR_DIR (or CURSOR_DB for a specific .vscdb file), ANTIGRAVITY_DIR, COPILOT_DIR / COPILOT_VSCODE_ROOT.

No install needed — run directly from GitHub:

npx github:naimjeem/agent-lens

Then open http://localhost:3456. The dashboard auto-detects which agents have data on your machine and disables tabs for the rest.

git clone https://github.com/naimjeem/agent-lens.git
cd agent-lens
npm install
cp .env.example .env

Edit .env to override data directories or pricing. All paths default to standard locations — usually no edits needed.

node server.js

Open http://localhost:3456.

Each agent has its own rate set. Defaults are sensible approximations; override per-agent:

Each prefix has four suffixes: _INPUT, _OUTPUT, _CACHE_READ, _CACHE_CREATE.

Default Claude rates match Bedrock cross-region inference (ap-southeast-2). For Anthropic API rates use RATE_INPUT=15, RATE_OUTPUT=75, RATE_CACHE_READ=1.5, RATE_CACHE_CREATE=18.75.

OpenCode also reports actual cost per assistant message in its DB; that value is preferred over the rate calculation when present.

• Cursor stores chat history in state.vscdb under Application Support/Cursor/User/globalStorage (macOS), ~/.config/Cursor/... (Linux), or %APPDATA%/Cursor/... (Windows). The dashboard reads the cursorDiskKV table directly: prompts, per-bubble token counts, tool calls (toolFormerData), and per-session model name from composerData. Override via CURSOR_DB (full path to state.vscdb) or CURSOR_DIR (directory containing it). Default rates are 0 because most Cursor users pay a flat subscription rather than per token; set RATE_CURSOR_* to estimate spend.
• Antigravity writes conversations as binary protobuf (*.pb) in ~/.gemini/antigravity/conversations. Without the .proto schema the dashboard reports session count and timestamps only.
• Kimi logs are sparse in the local wire.jsonl — turn boundaries and user inputs are reliable; token counts depend on whether your Kimi version emits TokenUsage/Usage events.
• GitHub Copilot chat sessions live under VS Code's User/workspaceStorage/<hash>/chatSessions/*.{json,jsonl}. The dashboard reads request prompts, response tool invocations, and the per-session model from inputState.selectedModel.metadata.name. Token counts are not stored in these files (Copilot bills via subscription server-side), so per-message tokens display as 0; messages, sessions, and tool calls are accurate. Adapter scans Code, Code - Insiders, VSCodium, and Cursor flavors. Override with COPILOT_DIR (a single User/ dir) or COPILOT_VSCODE_ROOT.

Maintained by naimjeem — <naimjeem.me>.