codexwatch 1.0.2

A small fix for the dashboard, mostly for Windows, and a tighter release build.

• The dashboard now serves its own files with fixed content types instead of the system's table. On Windows that table can label .js as plain text, and the browser then refuses to run the dashboard's script.
• The wheel and sdist are built with build and setuptools pinned by hash, so the build no longer fetches whatever setuptools is newest. The Docker image pins its base image by digest and installs nothing with pip.
• SECURITY.md links the private report form.

pipx upgrade codexwatch-cli      # or: pipx install codexwatch-cli

No changes to how anything is counted or shown. CHANGELOG.md

codexwatch 1.0.1

codexwatch's MCP server is now in the official MCP Registry as io.github.psychofict/codexwatch, so MCP clients that browse the registry can find and start it.

• codexwatch-cli works as a command as well as codexwatch, so uvx codexwatch-cli runs it without installing anything.
• The release workflow publishes to the registry after PyPI, signing in with GitHub's OIDC token rather than a stored secret.

pipx install codexwatch-cli      # or upgrade: pipx upgrade codexwatch-cli
codex mcp add codexwatch -- codexwatch mcp

No changes to how anything is counted or shown. CHANGELOG.md

codexwatch 1.0.0

The first release of codexwatch: see where your OpenAI Codex tokens, dollars and plan limits go, from the session logs Codex already writes on your machine.

pipx install codexwatch-cli
codexwatch              # the overview
codexwatch dashboard    # the web dashboard
codexwatch demo         # try it before you have logs

Every response is counted once. Rows are keyed on Codex's response id, so forked sessions, subagents, repeated token events and logs synced from another machine never add tokens twice.

Plan limits need no login. The 5-hour and weekly windows, burn rate and when a window fills come from the readings Codex writes into its logs after every response.

Quota forensics shows dollars of API-equivalent usage per percent of each window and day, and a flag on any rise your local logs cannot explain.

Also in this release: reports per day, project, model, session, tool and client; turn latency; prices from OpenAI's pricing pages (as of 2026-11-04) with long-context and Fast-tier rates, and unpriced models reported rather than counted as $0; a local web dashboard; a share card; SwiftBar, xbar, waybar, polybar, i3blocks, tmux and prompt output; a Linux tray; an MCP server for Codex and Claude Code; budgets with alerts; several machines in one history; and codexhop support with an opt-in autohop.

Python 3.9 or newer, no dependencies, Linux, macOS and Windows. The package is codexwatch-cli on PyPI because PyPI treats codexwatch as the same name as an unrelated project; the command is codexwatch.

Full list: CHANGELOG.md. How the numbers are made: docs/how-it-works.md.