AgentTelemetry

Live, local telemetry for your AI coding agents. AgentTelemetry reads the interaction logs your tools already write to your own machine and serves a clean, interactive dashboard — estimated spend, tokens, active time, cache efficiency, an Anthropic vs OpenAI vs Google provider comparison, an activity calendar, and breakdowns by model, provider, day, hour, weekday, tool, IDE, project and session — plus how much disk all these logs eat, and suggestions for spending less, drawn from your own numbers.

By default, log inventory, usage and the cache stay on this machine. Separate actions can share selected exports with paired devices, return requested analytics to an enabled local MCP client (which may forward tool results to its model provider), contact a provider after a confirmed Read once account lookup, or fetch update metadata after you click Check for updates. No account or API key is needed for the dashboard; it has no runtime dependencies beyond Python's standard library and a vendored copy of Chart.js.

Covers Claude Code · Claude Desktop · Codex · GitHub Copilot · Cursor · OpenCode · Hermes Agent · OpenClaw.

Download the app

Platform	Download
macOS 13+ · Apple Silicon	Download DMG
macOS 13+ · Intel	Download DMG
Windows 10/11 · x64	Download EXE installer

Python is bundled — no terminal setup needed. macOS: open the DMG, drag the app to Applications and open it. Windows: run the installer and open AgentTelemetry from Start. Open at Login is optional. Quit the app and wait for saving before installing an update; upgrades and uninstall preserve telemetry history.

If you have archived history in a source checkout, leave Open AgentTelemetry unchecked in Windows setup, open the app manually and Import existing data before starting monitoring. On macOS, opening the app manually also lets you import first.

These builds are unsigned/unnotarized. macOS may require System Settings → Privacy & Security → Open Anyway; Windows SmartScreen may require More info → Run anyway. All releases · SHA-256 checksums

---

Preview

Menu bar companion	Usage dashboard

Quick start

Install with Python 3.9+:

python -m pip install agenttelemetry-app
agenttelemetry

Or, with uv installed:

uvx agenttelemetry-app

Then open http://127.0.0.1:7878. The Python package starts the local browser dashboard; use the app downloads for the menu bar or system tray companion. See Python package details for its per-user data location.

To run from a source checkout with Python 3.8+:

git clone https://github.com/uttamdeb/agent-telemetry.git
cd agent-telemetry
python3 dashboard.py

Then open http://127.0.0.1:7878. The source checkout needs no package installation.

First run parses your local logs (can take ~30–60s if you have large Codex logs), writes a cache, and is instant thereafter. The page auto-refreshes every ~15s, so a session you're running right now shows up within seconds.

Options: python3 dashboard.py --port 9000 · --rebuild (ignore cache, full re-parse) · --interval 20 (background refresh seconds) · --data-dir /path/to/state (store the cache and optional device-sharing files outside the checkout). Without --data-dir, current checkout behavior is unchanged. Or ./run.sh [flags].

Editor extension

The AgentTelemetry editor extension brings local usage into VS Code compatible desktop editors, including Cursor and Antigravity IDE. Its Agent desk uses the app's icon and light/dark theme, with workspace totals, recorded main-agent/subagent relationships, and an in-memory usage baseline. Install the universal VSIX following the extension guide.

Recognizing subagent sessions

Sessions shows a full Subagent label, a branching icon, and a searchable session identifier for transcripts that the source identifies as delegated work. Use All sessions / Main agents / Subagents to focus the list; the choice survives reloads and shared links without changing the other dashboard totals. Session titles also open details from the keyboard.

Recorded delegation counts and Claude sidechain usage are explained in the drawer. Main agents shows sessions not marked as subagents; some tools and older logs do not record this marker. Parent links are not inferred. See behavior and limitations.

Recorded reasoning effort

Sessions shows Recorded effort below the model for Claude and Codex. Open a session for the recorded levels and their date/model scope. Changing the date, model or provider filter changes this label alongside the session figures. Missing metadata says Not recorded; mixed settings and partial coverage are explicit. These are recorded controls, not a measurement of reasoning quality or thinking tokens. See sources and limitations.

Subscription limits and account details

The bottom of Cost shows subscription limits for Codex and Claude Code when they have a session today on this device. Codex's locally recorded snapshots show reset times. Choose Refresh account & limits to read the installed Codex CLI's account, auth type, plan and current limits once. Claude's Refresh account reads its auth, provider and optional plan; percentage meters require the optional local status-line capture. Both actions ask before a lookup that may contact the provider.

Missing data stays unavailable. Older snapshots are stale; percentages are never estimated from tokens or combined across devices. Session details show recorded creator/account IDs with their attribution limits. Account metadata is private to the source device. See setup, research and limitations.

Local servers (macOS)

The Local servers tab and menu bar can list local project listeners, pin them, and save an explicit launch command for Start and Restart. Process actions ask for confirmation. They require a ready macOS backend; Windows and Linux show the support limitation. Commands stay on this machine and are never included in device sharing. Running a saved command gives it your account's file and network access. See behavior, privacy and limitations.

Install the menu bar / system tray app

Use the app downloads above for a ready-to-run installation. To build/install from a checkout, run python3 install.py on macOS or python install.py on Windows. The installer detects the OS, installs its matching native client for the current user and opens it. Use --no-launch to install without opening. Linux keeps the browser dashboard.

Platform	Native client	Source installation requirements
macOS 13+	SwiftUI/AppKit menu item, token/spend popover and embedded dashboard	Python to run the installer; Swift 5.8+ and a macOS SDK to build the app. The app then uses its bundled Python.
Windows 10/11	Branded tray icon and native summary popover matching the macOS layout; dashboard in your browser	Standard Python 3.8+ including pythonw.exe. No pip packages or separate GUI toolkit.

After installation, web Settings → Menu bar / system tray has a toggle to launch or quit the native app. Status reflects the app's heartbeat, not just a launch attempt. You can also open AgentTelemetry.app in ~/Applications or AgentTelemetry from the Windows Start menu. The app's Open at Login and Start monitoring when app opens options are separate and off by default. Opening from the installer or web toggle starts monitoring for that launch. No dashboard setting is required to open an app manually.

Both native popovers put a monitoring switch above today's token/spend cards, with expandable Details & settings, a dashboard button and always-visible refresh/quit controls. Windows follows the system light/dark or high-contrast theme, scales to the tray's monitor, and scrolls settings on smaller displays. Tab/Space/Enter operate the native controls; Escape or clicking elsewhere closes the panel. Icon mode uses the AgentTelemetry app icon; the numeric display remains available. Both have a text tooltip. Windows uses Segoe UI and opens the full dashboard in your browser; macOS uses its system font and embeds it.

Both clients reuse a dashboard already running on port 7878 and stop only a backend they started. A web/native app using a custom dashboard port is not supported by the native clients. Failure shows stale figures explicitly, and retries can recover. Import existing data before first app-owned monitoring if you need archived history: importing validates and copies the ledger without overwriting existing app data or changing the originals. For that flow, install with --no-launch, open the app manually and choose Import before Start monitoring. A default install attaches safely when the old dashboard is still running.

Browser windows, the macOS menu bar and the Windows tray share one refresh interval: 15 seconds (default), 1/5/15 minutes, or Manual. Change it in web Settings or the native app while monitoring. They sample a small shared clock each second and show figures from the same snapshot; a manual refresh in either app updates the others too, normally within a second plus request time. Manual continues listening for those explicit refreshes. The native figures are today's usage across all tools and connected devices; compare them with an unfiltered Today dashboard. Pause pauses only that browser's figures. Changing the interval does not restart the backend or reload the embedded dashboard. The server's --interval controls log parsing separately. Older attached backends without the sync API retain independent polling until upgraded and restarted.

Read the macOS guide and Windows guide for paths, controls and development checks. macOS downloads and source builds are ad-hoc signed; a Developer ID signed/notarized installer is not supplied. The backend remains stdlib-only and does not check for updates automatically. The macOS packaging step downloads a hash-pinned Python runtime explicitly.

Local app integrations can use GET /api/health for a small readiness response that contains no usage data, and GET /api/summary for today's token total and API-equivalent estimated spend. Both are read-only; they follow the dashboard's existing host and origin rules. GET /api/sync supplies the shared display clock; POST /api/sync sets its validated interval through the same JSON/same-origin CSRF guard. The data and summary HTTP endpoints retain one serialized snapshot per display tick, in memory only; the ledger is unaffected. GET /api/native reports local app status. POST /api/native requires the same CSRF guard as all writes. POST /api/shutdown additionally requires the owning Windows client's per-process control token; a normal standalone dashboard cannot be stopped by that route.

Local MCP access (optional)

AgentTelemetry provides eight read-only MCP tools for usage, comparisons, supported insights, session summaries and timelines, data health, the telemetry catalog and saved subscription limits. MCP access is off by default. Enable it in Settings and confirm each scope. The local bridge requires an already-running dashboard and returns data only when the configured MCP client requests it. A cloud-backed client may send tool results to its provider; review that client's data-handling settings before enabling access. Project/device names and usage from connected devices require their own opt-in scopes. The MCP bridge never refreshes accounts, returns prompt text, or edits client configuration. See MCP tools, setup and privacy limits.

--rebuild deletes the cache, and with it every session whose log has since been deleted from disk (see Storage). Copy .usage_cache.json somewhere safe first if you've cleaned up old logs.

Requirements: Python 3.8+ on macOS, Linux or Windows. On Windows run python dashboard.py (or run.cmd); on macOS/Linux python3 dashboard.py (or ./run.sh).

It works on anyone's machine because nothing is hardcoded — every location is derived at runtime from your own $HOME / %APPDATA% / %LOCALAPPDATA% / $XDG_*, the numbers are read live from your own logs on every refresh, and the disk figures come from your own drive. Two people running this see two completely different dashboards.

---

Everyone sees their own usage

This is the important part if you're sharing it: the dashboard has no bundled data. On each machine it scans that user's own logs (~/.claude, ~/.codex, ~/Library/Application Support/…, ~/.local/share/opencode, ~/.hermes, …) and builds a fresh .usage_cache.json locally. That cache is gitignored and never committed, so a clone starts empty and shows only the cloning user's numbers. (If you ever copy the folder instead of cloning, delete .usage_cache.json first — that file is your personal data.)

---

Data sources

Tool	Where it reads	Tokens
Claude Code	~/.claude/projects/**/*.jsonl	exact (in/out/cache read+write, 5m/1h tiers)
Claude Desktop (agent mode)	Claude/local-agent-mode-sessions/** under App Support / %APPDATA% / ~/.config	exact
Codex	~/.codex/sessions/**, ~/.codex/archived_sessions/**	exact (in/cached/out/reasoning); subagents are identified and labelled
GitHub Copilot	VS Code-family chatSessions/*.{json,jsonl}; SDK/CLI ~/.copilot/session-state/*/events.jsonl, enriched from VS Code agentSessionData/*/session.db	recorded request/turn tokens where available; SDK shutdown totals where safely dated. SDK gaps show a warning tooltip and unavailable/partial usage; older non-SDK chats retain text estimates. Premium multipliers are separate from estimated dollars.
Cursor (native AI)	Cursor/User/globalStorage/state.vscdb under App Support / %APPDATA% / ~/.config	partial — model, mode, timestamps, tool calls and AI-line stats are exact; tokens are on only ~2% of messages
OpenCode	~/.local/share/opencode/opencode.db, %LOCALAPPDATA%\opencode\opencode.db, ~/.opencode/opencode.db (or $OPENCODE_DATA_DIR)	V1/V2 SQLite; exact (in/out/reasoning/cache), prompts and tools; cost is read from OpenCode's own per-message value
Hermes Agent	~/.hermes/state.db (or $HERMES_HOME, %LOCALAPPDATA%\hermes)	exact (in/out/cache/reasoning, per model)
OpenClaw	~/.openclaw/agents/<agent>/agent/openclaw-agent.sqlite (or $OPENCLAW_STATE_DIR; also the older ~/.clawdbot, ~/.moltbot) and the older per-session sessions/*.jsonl	exact (in/out/cache read/cache write/reasoning, per response); its own logged cost is used only for a model with no verified price. Large events are stored zstd-compressed: read with the zstd command or Python 3.14+

A tool you don't use simply contributes nothing. Attribution is by tool, not by model — a Claude or GPT model used inside Copilot/Cursor/OpenCode/Hermes/OpenClaw counts under that tool, and the Models table lists each model × tool row separately.

Copilot's newer SDK-based agent sessions use a different store from older Copilot Chat. AgentTelemetry joins their event history to VS Code's saved per-turn usage using native session and event IDs. Custom endpoint models, including Cloudflare GLM 5.3, still count as Copilot. A warning beside a session explains missing usage, partial coverage, errors or interruptions. Missing SDK usage is never guessed from chat length or presented as a known zero. See Copilot SDK data coverage for sources and limits.

---

What you get

Eight views in a left-hand sidebar, light + dark theme, everything date-filterable, and the name of this machine at the top — so a screenshot always says which computer it came from.

The controls read as a sentence above every view — "Tokens from all tools over the last 30 days ‹ ›" — click a bold phrase to change it; the arrows step back and forward a period at a time. Inside them: the measure (Tokens · Cost · Messages · Time — tokens by default; every chart and ranking switches together), the period (today, the last 7–365 days, this week / month / quarter / year, last month, all time, or a custom range), and a single Filters panel (tool, provider, model, project, IDE / surface, and exact tokens only). Active filters show as removable chips. Every figure is compared with the equal-length period just before it, and every chart has a table view (the grid icon) so nothing is readable only by hovering.

Overview — one hero number (the chosen measure, tokens by default) with its change vs the previous period. In Tokens mode, hover over or keyboard-focus the total for a tooltip showing tokens excluding cache reads; Escape dismisses it. An estimated-spend sparkline sits below; Messages and Time keep spend in a tile. The per-day chart is stacked by tool, followed by supporting tiles, tool, model and project mix, highlights, an hour × weekday heatmap, and token mix per tool. The 12-month activity calendar and Monthly rhythm token chart let you click a day or month to select that period. Both keep their historical window while honoring tool, provider, model, project, IDE, device and exact-token filters; Monthly rhythm also has a table view with full counts. With another device connected, By device compares each computer's share of the selected measure (tokens, messages, estimated cost or active time), with a table view. It follows the same date and usage filters as the rest of Overview.

Cost — total with per-active-day, 30-day run rate, per session and per prompt · cumulative spend by tool against the previous period · cache hit rate and what caching saved you · spend by model, by project and by token type (what cache reads vs writes vs output actually cost you) · effective rate by model, as all tokens or per output (the one that's comparable across providers) · daily spend.

Models — provider cards (who made the model, independent of the tool that ran it) · sortable model × tool table with a ⚠ on any model missing a price · model timeline · provider share over time · provider × tool matrix · model efficiency (how often each model's edits land first time, retries per edit, tokens and cost per prompt).

Tools & agents — active time, tool calls per prompt, context amplification, subagent share, MCP and web calls, Copilot premium requests, Cursor's AI lines kept · top tool calls · calls by category · what your prompts were for (building, fixing, refactoring, testing, exploring, git, deploys… — each prompt classified by what the agent did: the files it edited and the commands it ran; Claude Code & Codex) · edits that landed first time (one-shot rate per tool, retries, and what the retried prompts cost) · where you work (IDE × tool) · MCP servers, with the tools used out of those each server offered · Skills · the full tool list.

Projects — ranked by the chosen measure, concentration stats, and a searchable table where clicking a row opens that project's sessions.

Sessions — real session titles, tool, project, model, tokens, cost, prompts, replies, tool calls, active time, cache % — search, sort, and click any row for a detail panel. Claude and Codex sessions also show completed compactions in the selected date range. Claude records manual/auto counts, context before → after, and duration in the detail panel. Codex checkpoints provide completion times; missing manual/auto triggers and context sizes show Not recorded. Newer Codex completion records can supply duration when their request/turn/thread identities and timing match exactly one checkpoint. Missing or ambiguous timing shows Unavailable. Codex compaction request usage, when logged, is already included in token/cost totals and is never counted again. Compaction-only days remain visible. Events survive log deletion once recorded; older ledger entries without event metadata say “Not recorded.” Context sizes are separate from billed tokens: missing compaction request usage/cost is not inferred. Events belong to the session and are not attributed to a model filter.

Use the pencil beside a session name (or Rename in its detail panel) to rename it in the original tool. A warning shows the old/new names and exact target before anything is saved. Close the session in its original tool before confirming, and reload that tool if needed. Token, cost, prompt, tool-call and compaction records stay intact. Previous title metadata is backed up privately in AgentTelemetryNative/session-title-backups/, separately from the usage ledger. This supports local Claude Code transcripts, Codex through its installed CLI’s official name API, and OpenCode’s SQLite session stores. Other formats, remote sessions and deleted logs cannot be edited. Recognized local Claude subagent transcripts instead offer Name in AgentTelemetry. This saves a private display name on this device without editing Claude's transcript, native subagent name, or parent session. The preview states this scope explicitly. These names live in AgentTelemetryNative/session-display-names.json, outside the usage ledger, summaries, and device sharing. Enter the source title again to remove a display-name override. See subagent name behavior. Tool/provider icons are bundled locally, including the OpenAI mark for Codex; no remote logo requests occur. Their sources and licenses are in static/tool-icons/README.md.

Optimize — suggestions derived from your own logs, ranked by what they'd save: sessions re-reading a very large context, thinking share, tool-heavy sessions that never delegated, cache written but never read, a costly model doing light work, what each Skill costs, edits that needed a retry, oversized CLAUDE.md / AGENTS.md, sessions that open heavier than your leanest ones, wasted file reads, installed skills and agents you never use, and MCP servers you never call, barely use, or load everywhere but use in one project. Nothing is shown unless your data supports it, and because the estimates overlap they are never summed.

Storage — see below.

Keyboard — 1/7/3/9/a ranges, m month-to-date, , / . previous / next period, Shift+T/C/M/A tokens / cost / messages / active time, [ sidebar, / search, t theme, r refresh.

Links — every filter is in the address bar, so a reload, a bookmark or a pasted link opens the same view. ?range= takes today, 7d, 30d, mtd… or custom with &from=2026-09-01&to=2026-09-15; then ?metric=, ?tools=, ?providers=, ?models=, ?projects=, ?ides=, ?devices= (repeat a param for each value), ?exact=1, ?q= for the search box, and #cost, #sessions… for the tab. For example /?range=7d&metric=cost&tools=claude&projects=sims#cost. Only what differs from the default is written, and changing a filter never adds a browser-history entry.

What changed? — Overview and Cost compare the selected range with the preceding equal-length range. Switch between projects, models, tools and (for tokens/cost) token types to see signed contributions, including decreases and categories that disappeared. Each breakdown reconciles independently to the total change. Click a category for current-period sessions or its previous value for earlier sessions. All usage filters apply; All time has no comparison. Missing prior observations and an unfinished current period are labeled. Contributions identify where usage changed, not a causal explanation. Cost remains an API-equivalent estimate.

Session timeline — the session drawer combines daily retained usage with recorded compaction markers and local Claude/Codex source events: model requests, tool calls, user turns and recorded subagent starts. No prompts or tool arguments are returned. The read is bounded to the first 64 MiB and 400 events in the selected date range; truncation is explicit. Date filters use local dates; model filters apply to events with model attribution. Archived/remote sessions and unsupported formats retain daily detail where available; missing timestamps or history are not reconstructed.

Backups and troubleshooting — Settings shows detected-source scan results, actionable parsing errors, ledger failures and models missing prices. An inactive tool is not reported as stale merely because it has no recent activity. Backups are private copies of the ledger in .usage-backups/ beside it (including with --data-dir), excluded from device sharing and installation payloads. A saved-ledger snapshot is kept on the first save each UTC day; updates and cache resets require a safety copy first. Retention keeps 7 daily snapshots and up to 5 each of manual, pre-restore and corrupt-file recovery copies. A failed daily backup does not stop normal ledger saving and is reported in Settings.

Restore shows stored-session counts and requires a one-use, five-minute confirmation ticket. Changed backups, disk contents or newly parsed usage invalidate the preview. The current ledger is preserved before replacement, including unreadable bytes as a non-restorable recovery copy. Backups cover aggregates, not original transcripts, device pairing settings or native preferences; local backups do not protect against loss of the drive. Use python tests/preview_features.py --port 8896 for an isolated preview with invented data.

Version and updates — the sidebar footer shows the running version. Check for updates asks GitHub (the only time the dashboard goes online, and only when you click); Update fast-forwards your checkout and restarts the server. It refuses if you have local edits or commits that aren't on GitHub. A copy that isn't a git clone has no update button.

---

Storage — what these logs cost you in disk

The tools you use write a lot to disk, and nothing else tells you how much. The Storage tab shows total footprint and per-tool bytes, a free-space gauge that warns when the drive is nearly full, storage accumulation over time, the largest individual log files, bytes per 1M tokens (which tool stores its history most expensively), AI data on disk AgentTelemetry does not analyse, and copy-paste cleanup commands generated for your own paths and your own shell (find on macOS/Linux, PowerShell on Windows). AgentTelemetry never deletes anything itself.

Deleting old logs does not shrink your analytics — AgentTelemetry keeps every session it has already parsed, so the cleanup is safe.

---

Cost notes (read this)

Costs are estimates computed as tokens × price — the tools store token counts, not dollars, so cost is always derived. Rates live in parser.py → PRICING as (input, output, cache_write_5m, cache_write_1h, cache_read) per 1M tokens; edit freely (recomputed on each request, no re-parse needed).

• Anthropic rates are current list prices (Opus 5.5 $4/$20, Opus 5 & 4.x $5/$25, Sonnet 5 & 5.5 $2/$10, Sonnet 4.x $3/$15, Haiku 4.5 $1/$5; cache write 1.25×/2× input for 5-min/1-hour, cache read 0.1× — 0.05× on Opus 5.5). OpenAI GPT-5.4/5.5/5.6/6 are verified from OpenAI docs; older/other models are estimates.
• Haiku 5.5 uses Anthropic's published rates: $0.10/$0.50 per million input/output tokens up to 100K prompt tokens, $0.50/$2.50 above 100K. The threshold includes cached input and applies to each request. Available Claude logs backfill this pricing detail on a normal restart; token and message counts stay unchanged. Older archives/peers and other tools without that detail use the short-prompt rate. Optimize model substitutions conservatively use the higher tier when the target request size is unknown.
• Price changes are date-aware. When a vendor cuts a price (OpenAI's GPT-5.6 cuts of 2026-07-30 and 2026-08-21), usage from before the change keeps the rate it had then — see parser.py → PRICE_HISTORY.
• These are API-equivalent values. If you're on a subscription (Claude Max/Pro, Codex, Copilot), you don't pay per token — the $ is "what this would cost at API rates."
• Cursor rarely logs token counts, and older Copilot chats have none, so their tokens (and thus $) are rough. Current Copilot builds log real per-request tokens, which are used. Copilot's honest metric is request count and its premium-request total (both shown); Cursor's is messages, tool calls and AI lines kept (also shown).
• Billing details that change the price are applied: prompt-cache writes at their own rate (Claude and GPT-5.6/GPT-6), Claude fast mode, US-only inference (1.1x) and web searches ($10/1K). A forked or subagent Codex thread's replay of its parent is skipped, so it isn't billed twice.
• A model with no price row reads as $0 — add it to PRICING (see below).
• Gemini (2.5, 3, 3.1 and 3.5 to 3.8: Pro, Flash, Flash-Lite, Computer Use) is priced when it runs through supported tools, at the Standard rate for prompts up to 200K tokens. Live, TTS, image and embedding models are not priced, and nor are the higher Pro rate for longer prompts or audio input. The Gemini CLI itself writes no token counts to read.
• Gemini 4 Argon is recognized across the supported tools and uses Google's announced introductory rates: $2 input, $10 output and $0.10 cached input per 1M tokens. Access is still limited to trusted testers, and Google has not listed a public API model ID. These are API-equivalent estimates for Argon usage found in logs. The later $4/$20 rate has no effective date yet and is not applied ahead of time; cache storage charges are not modeled.

Note on log retention

Some tools delete old logs. Claude Code prunes transcripts after cleanupPeriodDays (default 30); Codex keeps everything. The dashboard also keeps parsed sessions in its cache even after a tool deletes the on-disk log, so totals don't silently shrink once seen.

You can change Claude Code's retention window from the dashboard itself — the ⚙ button at the foot of the sidebar edits cleanupPeriodDays in your own ~/.claude/settings.json (leave it blank to fall back to the tool's default). The write is atomic and keeps a .bak; every other setting in the file is preserved untouched. It's the only file outside its own cache that the dashboard ever writes.

---

Extending it

• Add a model's price / fix an unknown model: edit PRICING (and, if needed, the model-name normalizer) in parser.py. Verify rates against the vendor's docs.
• Add a new tool: add its paths + a parse_* function in parser.py, wire discover() and update_file(), then add it to SRC/ORDER in static/core.js and give it a --t-<source> colour in static/app.css. The tool colours are a colour-blind-validated palette whose order is the safety mechanism — see AGENTS.md before changing them.

See AGENTS.md for a concise, agent-oriented guide (any coding agent can run and extend this from that file).

Files

dashboard.py (server + cache + cost + /api/storage) · parser.py (log parsers + pricing) · index.html (shell) · static/app.css · static/core.js · static/charts.js · static/views.js · chart.umd.min.js (vendored Chart.js) · AGENTS.md · run.sh.

Install it as an app (optional)

The dashboard can run as a standalone window in your Dock or taskbar. It's off by default — open the ⚙ menu and turn on Install as an app, then use your browser's Install / Add to Dock. That registers a service worker so the shell still opens when dashboard.py isn't running; your usage data is never cached, /api/ always hits the live server. Turning it off again unregisters the worker and clears its cache.

Your devices (optional)

Use your coding agents on more than one computer? You can see them all in one dashboard. It's off by default and works over your local network only. Open ⚙ → Your devices:

1. On the computer you want to read from, click Share this device…. It shows an address (like your-mac.local:7879) and a pairing code.
2. On the computer you're looking from, enter that address and code, then click Connect….
3. To see each one from the other, do both steps on both.

Once connected, a Devices filter appears and the header reads "…from all tools on all devices…". The sidebar device card also becomes a dropdown: choose All devices or one computer. Both controls share the same filter, which is preserved in the URL. The connected device's usage is fetched every minute, and a copy is kept in .peers/ so it still shows while that computer is asleep. Disconnecting deletes the copy.

What sharing sends: session titles (often the start of a prompt), project and branch names, log file paths, models, token counts, costs and times. Full prompts, replies and file contents are never sent. Sharing opens one read-only port (7879) that answers only to the pairing code. The dashboard itself stays on 127.0.0.1. The connection is not encrypted, so share only on a network you trust. The setup checks on the Optimize tab (instruction files, installed skills, MCP servers) and the Storage tab still describe only the computer you're looking at. .peers.json and .peers/ are gitignored.

If it can't connect, the error says which of these it is:

• The other device's firewall. On Windows, allow Python on private networks when asked, and set the Wi-Fi to Private (Settings → Network & internet). On macOS, allow incoming connections when asked.
• A VPN (like 1.1.1.1 with WARP) on either device can get in the way.
• A .local name that doesn't resolve. Use the IP address shown next to it instead.

Once connected, a device that moves to a new IP address is still found: every answer lists the other addresses it can be reached at, and those are tried when the saved one stops working.

Contributing

See CONTRIBUTING.md — it covers the setup, the hard rules (stdlib only, nothing hardcoded, never commit your cache), how to add a tool source or a model price, and how to test a change. AGENTS.md is the architecture guide. Please read the note on setting your git email before your first commit.

Found a security problem? See SECURITY.md — don't open a public issue.

License

MIT — see LICENSE.