A local output-speed monitor for Codex, built with Codex.
TokenPulse is a read-only tray app for Windows and Linux. Pick a model and a time window to see output speeds across conversations. It reads local Codex logs and keeps your history on your computer.
English · 简体中文
Synthetic demo data, not a model benchmark.
- One view per model. Combine conversations over 30 minutes, 1 day, or 7 days; inspect the weighted average, median, range, and sample count.
- Every measured output. Explore a scatter plot, hover for timing and tokens, and open the anonymous record. Keyboard navigation reaches overlapping points.
- Evidence behind the labels. Inspect global configuration, turn configuration tags, submitted settings, request values, and server-confirmed values separately. Unknown stays unknown.
- A quiet tray companion. Show an optional speed number or use a normal window when a tray is unavailable.
- Local history. Filter observations and export CSV. Light/dark themes, reduced effects, English, and Simplified Chinese are included.
TokenPulse reads existing local data. It does not change Codex settings, send model requests, read credential files, or upload telemetry. History excludes conversation text, tool arguments, and titles; identifiers are locally pseudonymized. Data and privacy →
Windows: download the portable ZIP from Releases, extract it, and open TokenPulse.exe. Keep the _internal folder beside the EXE. Python is included. The build is unsigned; licenses, library sources, and replacement instructions are included in the ZIP.
From source (Windows or Linux): you need Python 3.11+ and uv:
git clone https://github.com/zzccchen/token-pulse.git
cd token-pulse
uv sync --locked
uv run token-pulse --demoThe demo uses isolated synthetic data and needs no Codex account. To observe your local Codex activity, close the demo and run:
uv run token-pulseThe panel opens on startup. Closing it keeps the app in the tray when available; use the tray menu to quit. Without a tray, closing the last window exits.
uv run token-pulse --no-tray # Normal window, without a tray
uv run token-pulse --hidden # Start in tray; fall back to a window
uv run token-pulse --language en # Language override for this runThe source defaults to CODEX_HOME or ~/.codex. Change it in Settings or with --codex-home. See the user guide for storage, export, and troubleshooting.
Output TPS = output tokens / matching output-stream seconds
Model average = sum(valid output tokens) / sum(matching stream seconds)
These are completed-output estimates from client logs, not a live token counter or a controlled benchmark. Total output tokens can include reasoning and generated tool calls. Initial waiting and tool execution do not count as generation time.
When tokens and timing cannot be matched reliably, the record remains available but has no TPS. Ultra Fast and Fast have separate markers. Labels prefer request or submitted-setting evidence; a scoped configuration fallback is explicitly labeled configuration tag. None of these labels proves server-side priority. Measurement rules →
TokenPulse is an early 0.1.5 project. The original supported range is Codex CLI 0.153.4–0.160.0; newer 0.162/0.163 alpha session formats have received bounded local checks. Current adapters recognize scoped configuration tags in newer feedback logs without treating them as submissions, requests, or confirmation, and display ultrafast as a separate Ultra Fast mode. Internal log formats can change; see compatibility for the exact validation scope.
| Environment | Validation scope |
|---|---|
| Windows 11 x64 | Native tray and window behavior validated locally |
| Ubuntu 24.04 under WSL2 | Wayland window and no-tray fallback validated locally |
| Full Linux desktops | GNOME/KDE/Xfce tray behavior still needs testing |
Missing or rotated logs can leave gaps. Final server-confirmed service tiers are not currently extracted. Remote hosts, other clients, cost tracking, and changing Codex settings are outside the current scope.
Windows portable builds include dependency notices, corresponding Qt/PySide sources, and SHA-256 checksums. Clean-machine Windows testing remains pending. There is no PyPI release or Linux binary yet. Compatibility · Build instructions
Useful first contributions include Linux desktop validation, synthetic format reproductions, translation improvements, and accessibility fixes. Please do not upload real conversation logs.
- Report a bug
- Suggest an improvement
- Contribution guide · Development · Architecture
- Roadmap · Security reporting
Found a problem? Open an issue. English and Chinese are both welcome.
MIT for TokenPulse's own code. Dependencies retain their licenses; see third-party notices.
An independent community project. Not affiliated with or endorsed by OpenAI.