I wanted one number for “what have I burned on models this week” without opening a dashboard. What I ended up with is a five line shell script that two separate sources answer from:
#!/bin/bash
LLM_COST_RAW=$(tokscale --week --json | jq .totalCost)
LLM_COST=$(printf "%.2f" "$LLM_COST_RAW")
GH_USAGE=$(curl --silent http://localhost:4141/usage | jq .quota_snapshots.premium_interactions.quota_remaining)
echo "$ $LLM_COST GH $GH_USAGE"
Output looks like $ 2.26 GH 412. Two things in there are worth explaining, because neither is obvious.
tokscale is a CLI that reads the local session logs of coding agents (Claude Code, Codex, OpenCode, Hermes, and friends), maps token counts onto current model prices, and reports the total. --week scopes it to the rolling seven day window and --json makes it machine readable, so jq .totalCost gives me the dollar figure and printf "%.2f" rounds it before it hits the terminal.
The 4141 endpoint is not a public service. It is a local proxy that fronts a Copilot-style API, and /usage returns a quota snapshot with premium_interactions.quota_remaining inside it. That line exists so I can see model spend and subscription headroom next to each other, since they come out of different budgets and I want both in the same glance.
Caveats, honestly
The first half is portable. Install tokscale, install jq, done. The second half is specific to my box: if you do not run something on port 4141 that speaks that API shape, the curl returns nothing, jq on empty input prints null, and the script quietly prints GH with no number after it. Drop line 4 and the trailing GH $GH_USAGE if that is you.
On the day I wrote this the proxy was down, so the script was printing a dollar figure and an empty GH field. That is the accurate state of it. I am posting it as it stands rather than tidying it up, because the pattern is the useful part: a cheap weekly scoreboard built out of two local data sources that were never designed to be combined.