Usage

Usage answers two questions: how many tokens have my agents spent, and what are Yardsort and its agents using of this machine right now? Open it with Usage at the foot of the sidebar, beside Settings, or from the command palette (Mod+K, then Usage). It takes over the center panel the way a workflow does; × in its header closes it, and so does selecting a workspace.

It has two tabs, Token usage and Machine resources. Usage remembers the tab you last looked at until you quit.

Nothing here is sent anywhere. Token figures come from files the agents keep on this machine, and machine figures from the operating system.

Token usage

Where the figures come from

Each agent keeps a log of its own conversations, with the tokens of every model call. Yardsort reads those files where they are and never changes them:

Agent	Where its logs are	What one entry is
Claude Code	projects/ in CLAUDE_CONFIG_DIR, else in ~/.claude (and ~/.config/claude, where some installs keep it)	One API call
Codex	sessions/ and archived_sessions/ in CODEX_HOME, else in ~/.codex	One model response
Grok	sessions/ in GROK_HOME, else in ~/.grok	One turn, split by model where it was

The variables are read from the same environment your agents are started in, so the folders are the ones they actually use. A line under the tabs lists each agent found, where, and how many log files it has.

This counts every conversation those agents had on this machine — started in Yardsort or in a terminal of your own. It knows nothing of other machines. Other agents (OpenCode, OMP, pi, Cursor) are not counted yet.

Claude Code writes a reply once per part of it, and a resumed conversation copies earlier replies into its new file, so each call is counted once by its own ids. Codex releases before token_usage_record only wrote running totals; those are counted by how much each total grew.

The logs are read when the tab opens, when you change the range, and when you press ↻. Files that have not changed since the last read are not read again, and files last written before the range starts are not opened at all.

Cost

Cost is an estimate: what the same tokens would have cost billed per token at each vendor's published API rate. It is marked with * for that reason. If you use a subscription — Claude Max, ChatGPT Pro, SuperGrok — you are not billed this way, and this is not what you pay; it is a way to compare agents, models and workspaces with one unit.

The rates are built into this version of Yardsort, from Anthropic's and OpenAI's price lists: input, output and cache reads per model, and for Claude, cache writes at 1.25× input for the five-minute cache and 2× for the hour-long one. A model whose price is not known — every Grok model, for now, and anything newer than this release — has its tokens counted and its cost left out, shown as —, and named under the total as Not priced. A near match is never used: gpt-5.4-pro is not priced as gpt-5.4.

What is on the tab

• Plan limits. Codex writes its plan's rate limits into its log each time its vendor tells it: the plan, how much of each window is used, and when the window resets. They are shown as Codex last heard them — as of 3h ago — and a window that has reset since says so. Claude Code and Grok do not write theirs down, so they are not shown.
• Range and unit. 7d, 30d and 90d cover the last seven, thirty or ninety days, today included, in your time zone. Cost and Tokens switch every figure on the tab between dollars and tokens.
• The total and each agent's share, beside a bar for each day with the agents stacked in it. Hover over a day for each agent's figure and the day's total.
• Processed tokens — everything: input of every kind, and output. Cached input is input read from a prompt cache, with its share of all input. Cache writes is input written to one (Claude only; the others do not charge for it separately). Uncached input is the rest. Output includes reasoning. Cache savings is what reading from the cache saved against paying full input price for the same tokens.
• By model, most expensive first: cost, share of the cost and tokens. All → lists every model.
• By workspace. A call is put in the Yardsort workspace whose folder it ran in — the deepest one, so a worktree inside a project's folder is its own. Click one to open it. A folder Yardsort does not know is listed by its name; hover for where it is.

When no logs are found at all, the tab says where it looked.

Machine resources

While this tab is open, Yardsort samples every two seconds; while it is closed, or the window is hidden, nothing is sampled. ↻ samples at once.

• Yardsort CPU is a share of the whole machine — every core counted, so it never passes 100 %. Yardsort memory is resident memory. Both include the app, its terminal host and every terminal it runs, with everything those started.
• Share of RAM, System memory (in use, and as a share of all of it), CPU cores, and Load (1 m) — the one-minute load average. Windows has no load average; it shows System CPU instead.
• The bar beneath splits memory into Yardsort, other apps and free.
• Two charts show Yardsort's CPU and memory over the last five minutes it was looked at. Hover for a moment's figures.

The table breaks Yardsort down:

• Yardsort app — Main is the app's own process; Webview is the processes below it that draw the window; Terminal host is the background process that keeps terminals alive, without the terminals.
• Then each project with something running, each workspace in it, and each terminal: a conversation by its title, else the agent's name, a shell as Shell and a run command as Run command. A terminal's figures are its whole process tree — the agent and every program it started, such as a test run or a language server — so nothing is counted twice. Hover a terminal for how many processes that is. The arrows collapse a project or a workspace; click a workspace's name to open it.
• Terminals in no workspace are listed last, under Other terminals.

Sort switches the order between memory and CPU.

Memory is summed over processes, as every task manager does, so pages that several processes share are counted once for each. On macOS the processes that draw web content belong to the system rather than to Yardsort, so Webview there shows little or nothing.

Hiding Usage from the sidebar

Show Usage in the sidebar, at the right of the Usage header, takes the button off the foot of the sidebar, or puts it back. The command palette offers Usage either way. The choice is kept in settings.toml:

[usage]
show_in_sidebar = false