Tokentelemetry

The pitch is unusually honest: TokenTelemetry doesn't ask you to instrument anything. It scans the log directories your agents already populate — ~/.claude/projects/ for Claude Code, ~/.codex/sessions/ for Codex, ~/.gemini/ for Gemini CLI and Antigravity, ~/.cursor/ and workspaceStorage for Cursor, VS Code chatSessions for Copilot — and serves a read-only UI at localhost:13000. That read-only stance has real consequences worth understanding before you install. The agents must already be installed and logging on the same machine; there is no remote or server-side collection, so a teammate's spend is invisible to you unless you run it on their box. Nothing is written back to agent files, which means you can't use it as a control plane — this is observability, not orchestration. Where it earns its place is day-two questions: why did last week's spend spike (cost anomaly detection flags silent reasoning-token waste), which model is cheaper for this project (per-agent and per-model breakdowns with date-range filters), and what did that multi-step run actually do (replayable traces with tool calls, subagent delegation, plan-mode output, and an honest label on encrypted reasoning rather than pretending to decode it). Analytics adds cache-efficiency metrics and a Delegation & Ecosystem section covering subagent spend, skills used, and MCP servers — the last one being genuinely rare in local tooling. Summarization is one-click, into either a narrative or a structured What/Tools/Why/Next brief, cached by content hash, and requires configuring your own LLM. Projects gives one card per working directory with a 365-day heatmap and streaks, which is more useful than it sounds for spotting which repo is burning tokens. The Hermes surface is the standout: Hermes runs across CLI, Telegram, Discord, Slack, Feishu, DingTalk, cron and webhook — 38 source platforms — and TokenTelemetry is the only tool in this comparison that observes them as a single agent, with per-API-call latency and cache-hit percentage parsed from agent.log, 90 loaded skills with platform conditions, and MEMORY.md / USER.md char-limit progress bars. A separate one-command plugin ($ hermes plugins install VasiHemanth/tokentelemetry-hermes-plugin) puts a TokenTelemetry tab inside Hermes Dashboard on port 9119, though it is a launcher, not the engine — TT still has to be running on 13000. The constraints are real. It needs Node 18+ and Python 3.9+ locally. Two outbound features are on by default — the hourly GitHub update check and anonymous content-free product telemetry — and while both are one-click off in Settings → Usage & privacy or killable with DO_NOT_TRACK=1, organizations that block unsanctioned local services or curl-pipe-bash installers will need to clone and run it instead. There is no shared cloud dashboard, no multi-user seats, and no vendor SLA. The MIT license with 180 public commits and an allowlist test in the telemetry pipeline is the mitigating factor: you can read the