What a flat-rate plan actually buys, measured.

One developer’s last two weeks

94% of prompt tokens were read from cache. Plan windows exhausted in the latest week:

• claude-code 5h window, reached 2026-09-30 03:40 UTC (reset 2026-09-30 07:20 UTC) · also hit per model: claude-haiku-4-5, claude-opus-5, claude-sonnet-5-5, claude-fable-5-1

One developer's own recorder, seven-day windows ending at capture. Usage priced at each model's list rate on the capture date; the developer paid flat-rate subscriptions, not these amounts. Captured 2026-10-01 22:52 UTC.

Observed usage per meter point

How much local work was recorded while a shared subscription meter moved by one percentage point. Calls outside the recorder can consume the same meter. These observations do not measure the full allowance or what a subscription costs.

Checking shared meter evidence…

Only bound subscriptions, unchanged reset windows, known token categories, and nonoverlapping observed installations qualify. Unknown bindings and scoped meters are retained privately. Owner and install thresholds reduce duplicate influence; they do not prove that subscriptions are independent. Five-hour and weekly meters can cover the same calls: do not add them together. View the aggregate API.

What it does

• Runs a local proxy on 127.0.0.1:4141 for Claude Code and Codex: every attempt is recorded live, with retries, cancellations and the rate-limit headers the provider returns.
• Warns a session, or asks before it continues, when a plan window is on pace to run out; tokenpylon guard check gives the same answer to scripts as an exit code.
• Shows spend at list for the session, today and the last hour in the Claude Code status line, and a weekly report with limits reached (tokenpylon weekly, also as CSV).
• Reads local history from Amp, Antigravity, Codebuff, GitHub Copilot CLI, Droid, Gemini CLI, Goose, Grok CLI, Hermes, Kilo, Kimi CLI, OpenClaw, OpenCode, Pi, Qwen Code, Z Code.
• Has a local dashboard at http://127.0.0.1:4141/ and an MCP server (tokenpylon mcp) any agent can query.

Install

Private beta: you need access to the TokenPylon GitHub repository. Install GitHub CLI and run gh auth login first; downloads use your existing GitHub login.

macOS or Linux:

curl -fsSL https://tokenpylon.com/install.sh | sh

Windows (PowerShell; the Windows build is not code-signed, so SmartScreen may warn):

& ([scriptblock]::Create((irm https://tokenpylon.com/install.ps1)))

The installer checks each download against its SHA-256 checksum. macOS builds are signed and notarized. Then:

tokenpylon setup                # checks history, connections and sharing
tokenpylon connect claude-code  # route Claude Code through the recorder (also: codex, shell)
tokenpylon up                   # start it at login

To add your readings to the shared measurements, create an account, make a key under settings, and paste it into tokenpylon auth set --stdin.

What it shares

Sharing is on by default. Each call sends its metadata: model asked for and served, provider, token counts by kind, cost, timing, status and the provider’s request id. Quota uploads also include the harness, reported plan, meter window and scope, used percentage, reset time (or unknown), observation time, and a captured unit binding when proven. Messages, prompts, provider keys and project names are excluded from ordinary sharing. tokenpylon sharing show prints the exact fields; tokenpylon sharing off or TOKENPYLON_SHARE=0 stops sharing, and everything local keeps working. See the privacy policy.