AgentTelemetry 2.1.0

AgentTelemetry 2.1.0

AgentTelemetry 2.1.0 brings the original native app release forward with the reviewed
dashboard, MCP, editor, telemetry, and native-client improvements made since 2.0.0.
All components in this release use version 2.1.0.

Downloads and listings

Product	Install or download
macOS 13+ · Apple Silicon	Download DMG
macOS 13+ · Intel	Download DMG
Windows 10/11 · x64	Download EXE installer
VS Code	Visual Studio Marketplace
Cursor, Antigravity and compatible editors	Open VSX · Universal VSIX
Python dashboard and MCP bridge	PyPI 2.1.0 · pip install agenttelemetry-app · uvx agenttelemetry-app
Local MCP setup and privacy	MCP guide
Product and install guide	AgentTelemetry landing page
Integrity	SHA-256 checksums

The native app includes Python and runs the same local dashboard on macOS and Windows.
There is no native Linux tray app in 2.1.0; Linux users can run the browser dashboard
from the Python package or source checkout. The MCP
bridge is included in the Python package and native/editor backend; it is read-only, uses
stdio, requires an already-running dashboard, and remains off until enabled in Settings.
It is a component of AgentTelemetry 2.1.0, not a separate marketplace download.

What's included

• Local dashboard and native clients: macOS menu bar and Windows system tray apps,
  recent-session activity, local-server controls, shared refresh behavior, release notes,
  and the same local ledger and backend.
• Session and delegation detail: clearer main-agent and subagent views, recorded
  parent navigation where the source provides it, Claude and Codex reasoning-effort labels,
  source-recorded model and tool activity, compactions, and timeline filtering.
• Copilot telemetry: workspace attribution, named SDK child sessions, improved usage
  accounting and compaction labels, and explicit reporting limits for partial SDK totals.
• Agent desk extension: workspace-first usage, linked sessions and delegation, live
  recorded-usage counter, native activity sidebar, and dashboard access. The same
  cross-platform VSIX is listed separately in Microsoft Marketplace and Open VSX.
• Read-only local MCP: eight analytics tools for usage, comparisons, insights, sessions,
  timelines, data health, dimensions and saved subscription-limit observations. Access and
  optional data scopes require separate consent.
• History and project continuity: safer project relocation handling, ledger backups and
  recovery, and preserved archived history during supported metadata upgrades.

Cost figures are estimates at published API list prices, not provider bills or subscription
charges. Coverage and detail depend on what each coding tool records. The macOS DMGs and
Windows installer are unsigned/unnotarized; review the linked release installation notes
and checksums before installing.

AgentTelemetry 2.0.0

Download AgentTelemetry 2.0.0

This refreshed 2.0.0 package includes the original native-app release and all reviewed application changes through October 18, 2026.

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

Python is bundled. Open the macOS DMG and drag AgentTelemetry to Applications, or run the Windows installer and open AgentTelemetry from Start. Open at Login remains optional.

These downloads are unsigned/unnotarized (macOS uses ad-hoc signing). macOS may require System Settings → Privacy & Security → Open Anyway; Windows SmartScreen may require More info → Run anyway. SHA-256 checksums.

Included in this consolidated release

• Native apps: macOS menu-bar and Windows tray clients, Today tokens and estimated cost, web launch/quit controls, shared refresh cadence, stale/retry handling and graceful ledger saves.
• Dashboard: token-first Overview, tool icons, device filtering, accessible chart tables, retained links/filters and improved sidebar layout.
• Sessions and delegation: All sessions / Main agents / Subagents filters, recorded parent navigation, Claude child-session associations, scoped model labels and source-recorded reasoning effort. Missing relationships and effort remain unavailable rather than guessed.
• Session names: confirmed original-tool title editing for supported Claude, Codex and OpenCode sources, with private metadata backups; local display names for supported Claude subagents.
• Compactions and parsing: Claude/Codex compaction events and recorded Codex durations, OpenCode V2 history, Copilot SDK/Agent Host recovery, Windows 11 labels and Haiku 5.5 request-tier pricing.
• Usage explanations: current-versus-previous comparisons and contribution breakdowns, including decreases and zero baselines.
• History protection: scheduled/manual private ledger backups, confirmed preview-and-restore, source health diagnostics and bounded on-demand session timelines.
• Local servers: macOS and Windows listener discovery and confirmed lifecycle controls, with saved commands kept separate from telemetry.
• Account snapshots: explicitly requested read-only CLI lookups and saved quota observations, with freshness and unbound-identity labels. API-price estimates are not bills or subscription allowances.
• Optional MCP: eight read-only analytics tools, disabled by default. Enable in dashboard Settings, review the cloud-client warning and copy the generated client configuration. Project/device display names and paired-device data require separate consent. Setup and privacy guide.
• Desktop editors: the AgentTelemetry Agent desk, workspace usage brief, delegation outline and activity sidebar in Visual Studio Marketplace. The universal extension also supports compatible desktop Cursor and Antigravity hosts via VSIX.

Build provenance and upgrading

Application/native version stays 2.0.0; the Marketplace extension update is 2.0.1 and the PyPI packaging revision is 2.0.0.post1.

The original public v2.0.0 tag is unchanged. These refreshed download assets were built from 3b094bb; see BUILD-MANIFEST.json. Use the attached AgentTelemetry-2.0.0-refreshed-source.zip for matching current source. GitHub's automatic “Source code” archives still represent the original tag.

Quit the native app before replacing it. Back up your ledger before upgrading; normal upgrades retain archived history and do not need a rebuild. Backend/frontend, MCP client compatibility and editor checks pass; both macOS DMGs and the Windows installer undergo native/runtime verification before publication.

v1.7.1

AgentTelemetry v1.7.1 adds device distribution to Overview and corrects verified ledger, session accounting, pricing and filter issues.

• By device: compare each connected computer's tokens, messages, estimated cost or active time using the global measure and filters. Includes a table view and hides on single-device installs.
• Durable history: failed incremental parses roll back; failed saves retry; unreadable ledger files are preserved. Restarts wait for a successful ledger save.
• Accounting: multi-session stores retain project/day/model attribution, mixed-model and cache-tier costs reconcile, and overlapping opencode migration messages count once per device. Copilot exact-only retains recorded counts and request ordering produces deterministic time estimates.
• Pricing and filters: dated aliases and fast/US rates agree across Python and JavaScript, secondary-model sessions follow filters, and unsupported ratios are clearly limited. Optimize removes an unsupported cache-saving amount and scopes MCP evidence by date, tool and device, including rare calls.
• Local HTTP boundary: dashboard reads validate Host/Origin; refresh uses guarded JSON POST.

Validation: 24 isolated stdlib tests, frontend regression checks, JavaScript syntax checks, Python 3.8 grammar checks and 64 browser render combinations, plus device-chart/filter/table/mobile checks. A copied-ledger migration preserved 253 archived entries and reconciled usage by device, source and date.

The cache schema is now 53. Normal startup migrates live logs while keeping archived history. Version 52 (v1.7.0) device exports remain readable; missing detail in older archived/remote stores is labelled rather than reconstructed. No runtime dependencies or vendor price changes were introduced.

v1.7.0

Gemini 4 Argon usage is now recognized and priced across supported tools. Live usage refreshes also retain the last valid history when a full reparse fails, and detect Cursor/Hermes updates that exist only in SQLite's write-ahead log.

Gemini 4 Argon

• Normalize Argon display names and raw log spellings through the shared Gemini normalizer, including provider prefixes and preview suffixes.
• Use Google's announced introductory API-equivalent rates: $2 input, $10 output and $0.10 cached input per million tokens. The later $4/$20 rates have no effective date and are not applied in advance.
• Access remains limited to trusted testers; Google has not listed a public API model ID. Support was verified with synthetic log fixtures, since no live Argon usage was available locally. Audio, image and Live variants are not assigned text-model prices. Gemini CLI remains unsupported because its local bookkeeping lacks usable usage data.

Fixes

• Correct the Copilot pricing note: newer logs provide recorded request token counts; older or unfinished requests use estimates from visible text.
• Report full-parser failures during background refreshes, retain the prior aggregate/signature, and retry on the next refresh. Close database handles on failures.
• Detect WAL-only changes in Cursor and Hermes stores, instead of waiting for a database checkpoint.
• Serialize payload construction and cache actions with refreshes to prevent concurrent mutation and lost cache updates. Snapshot Storage metadata before scanning the filesystem.
• Deduplicate conversations within each device, recognize Windows path separators on other operating systems, and include cache-write tokens when choosing the fullest copy.
• Skip files that disappear between discovery and parsing rather than saving a null ledger entry.
• Document normal upgrades that preserve archived history, and add a stdlib regression suite using temporary stores.

Upgrade

The cache version is now 52. A normal restart reparses live logs while retaining archived ledger entries whose source logs no longer exist. Back up .usage_cache.json before upgrading; do not use --rebuild for a routine upgrade because it discards archived history.

Update connected devices to v1.7.0 as well. Devices on v1.6.x use cache version 51 and cannot sync with this release until updated. Saved peer mirrors remain on disk.

Validation

• python3 -m unittest discover -s tests -v: all 9 regression tests passed, covering model rates, token fallbacks, WAL-only writes, parser failures, archive migration, deduplication and refresh concurrency.
• All 59 existing model-price rows and their normalizations were unchanged; a frozen existing ledger produced identical usage, costs, sessions and breakdowns.
• Read-only snapshots of representative live SQLite stores matched the previous parser's successful output.
• All 8 views rendered without errors across light/dark themes and today/30-day ranges (32 browser checks). Argon filtering/cost display and write-endpoint CSRF rejection also passed.
• Python 3.8 syntax compatibility and git diff --check passed.

v1.6.6

Prices for GPT-6.1 Sol and every Gemini text model, and three OpenAI prices corrected.

New

• GPT-6.1 Sol is priced. It had no price, so a session running it showed its tokens but $0. It costs $2 in and $10 out like GPT-6 Sol, with cached input at $0.10 instead of $0.20.
• Gemini is priced. Gemini 2.5 (Pro, Flash, Flash-Lite, Computer Use), 3 Flash, 3.1 (Pro, Flash-Lite), 3.5 (Flash, Flash-Lite) and 3.6, 3.7 and 3.8 Flash now have prices, so Gemini used through Copilot, Cursor or opencode no longer reads $0. Every tool shows the same name, for example "Gemini 3.1 Pro".
• Live, voice, image and embedding Gemini models are not priced. They bill audio, image and video at other rates. Pro prompts over 200K tokens and audio input are billed at the standard rate too, so those are a little low.
• 3.6 to 3.8 Flash use Google's promo rate, which runs through 2026-12-31 and doubles on 2027-01-01.

Fixed

• GPT-5.2 and GPT-5.3 Codex were too cheap. They were priced like GPT-5.1 ($1.25/$10). The rate is $1.75/$14, cached $0.175, so their past cost goes up.
• GPT-5 Nano had no price and cost $0. It is now $0.05 in, $0.40 out.
• Anthropic's whole price table was checked against Anthropic's page, and nothing needed changing.

Upgrading

• Update each computer you look at the dashboard on. Click Update in the sidebar footer, or run git pull, then restart so the new prices load.
• No re-parse is needed. Prices apply when the page is built, and the cache format hasn't changed.
• Devices on any v1.6 release can still connect to each other.

v1.6.5 — Filters are saved in the address bar

Every filter is now saved in the address bar, so a reload, a bookmark or a pasted link opens the same view.

New

• Filters live in the URL. The period (a preset or a custom range), the measure, tools, providers, models, projects, IDEs, devices, "Exact tokens only", the search box and the "Per output" switch on the Cost tab are all written to the link as you change them. Before, they reset whenever the page reloaded, including the automatic reload after an update. For example: /?range=7d&metric=cost&tools=claude&projects=sims#cost.
• Only what you've changed is written. A dashboard with nothing filtered keeps a clean link, and changing a filter doesn't add a browser-history entry, so Back doesn't step through them.
• Links are checked before they're used. Dates, tool names, the period and the measure are validated, and project, model and search text is only ever shown as plain text. A value that matches nothing shows as a chip you can remove.
• Links you already have keep working: ?range=today, ?metric=cost and ?theme=.

Upgrading

• Update each computer you look at the dashboard on. Click Update in the sidebar footer, or run git pull.
• No re-parse is needed. The cache format hasn't changed.
• Devices on any v1.6 release can still connect to each other.

v1.6.4 — Claude Sonnet 5.5 is priced

Claude Sonnet 5.5 now has a price, and a session that switched models is named after the one it used most.

Fixed

• Claude Sonnet 5.5 showed no cost. Its tokens were counted, but with no price on file its estimated cost was $0 wherever it appeared, so every total left it out. It's now priced at Anthropic's published rate: $2 per million input tokens and $10 per million output tokens, $0.20 to read the cache, and $2.50 or $4 to write it for 5 minutes or an hour. That's the same as Sonnet 5. The rate applies to Sonnet 5.5 usage you've already logged, so earlier days fill in too.
• A session that switched models was named after the model it used last. The Sessions table's Model column, and the order of the models listed in a session's details, followed whichever model the newest part of the log used. A long Opus session that ended with a few Sonnet 5.5 replies read "Claude Sonnet 5.5". It now names the model that did most of the work. Tokens and costs were never affected, since they're priced per day and per model.

Upgrading

• Update every computer you look at the dashboard on. Each one prices the usage it shows with its own price list. Click Update in the sidebar footer, or run git pull.
• No re-parse is needed. The cache format hasn't changed. A session nobody has used since keeps its old model name until its next message.
• Devices on any v1.6 release can still connect to each other.

v1.6.3 — Connected devices stay in sync

A fix for Your devices: a connected device now keeps up with usage as it happens.

Fixed

• Usage from a session that's still running now reaches your other device. Before, the device you were looking from kept showing the snapshot from when it connected. For example, it stayed at 16.1M tokens while the other computer was already past 22.3M. Only a new session, or restarting the dashboard on the sharing computer, sent fresh numbers. The connected device's usage now refreshes every minute, as intended, and a minute with no new usage still sends nothing.
• A long session's progress is now saved to the cache. It's written at most every 5 minutes, so restarting the dashboard doesn't have to re-read as much.

Upgrading

• Update the computer that's sharing. That's where this fix takes effect. Click Update in the sidebar footer, or run git pull. Updating both is best.
• No re-parse is needed. The cache format hasn't changed.
• Devices on any v1.6 release can still connect to each other.

v1.6.2 — The version label keeps up

A small fix to the version shown at the bottom of the sidebar.

Fixed

• Check for updates now shows the right version when you're already up to date. A release can be tagged on a commit you've already downloaded. The check then correctly says "Up to date", but the sidebar kept the version it read at startup. After updating to the v1.6.1 fix, for example, it still said v1.6.0 +1. The check now fetches the new tag and the label updates straight away, with no restart.

Upgrading

• Click Update in the sidebar footer on each device, or run git pull.
• No re-parse is needed. The cache format hasn't changed.
• Devices on v1.6.0, v1.6.1 and v1.6.2 can all connect to each other.

v1.6.1 — Devices connect more reliably

Fixes for Your devices, found while connecting a Mac and a Windows PC.

Fixed

• The address under Share this device is now one your other device can reach. With a VPN like 1.1.1.1 (WARP) turned on, it used to show the VPN's own address (like 172.16.0.2), which nothing else on your Wi-Fi can reach, so connecting to it always failed. It now shows your Wi-Fi or Ethernet address.
• A connection survives a new IP address. Each device tells the one reading it every address it can be reached at. If the saved one stops working, the others are tried. This covers your router handing out a new IP, or a .local name that didn't resolve. The device card shows which address answered.
• A failed connection says why, within about 5 seconds (it used to take 20). It tells you whether the name couldn't be found (use the IP instead), sharing is off on the other device, or nothing answered at all (usually a firewall or VPN).
• The firewall tip matches your OS. On Windows, allow Python on private networks when asked, and set the Wi-Fi to Private in Settings → Network & internet.
• Addresses no longer break across two lines.

Upgrading

• Click Update in the sidebar footer on each device, or run git pull.
• No re-parse is needed. The cache format hasn't changed.
• v1.6.0 and v1.6.1 can still connect to each other. Update both to get every fix: the address shown comes from the sharing device, and the fallback needs both.