Token Receipt | Your Agent Has Expenses

Codex and Claude Code support<br>Both agents get the same skill-first flow. The skill calls the local runtime, reads the structured analysis, and then lets Codex or Claude Code write the final response using the session you are already in, instead of a separate API key.<br>Quickest<br>/token-receipt

Or be specific<br>/token-receipt Generate a receipt for my last 30 days of Codex, Claude Code, Kiro CLI, and Cursor usage.

What is Token Receipt?<br>Token Receipt is a screenshot-first usage artifact for coding agents.<br>It reads the session logs that Codex, Claude Code, and Kiro CLI already write locally, plus experimental local Cursor workspace metadata and request traces, turns those into structured usage facts, and lays them out as an itemized bill.<br>The result feels personal because it reflects your own habits: repeated file reads, subagent sprawl, context bloat, and every other expensive little ritual.<br>Reads local logs from Codex, Claude Code, Kiro CLI, and experimental Cursor local sessions<br>Kiro spend uses local credit usage instead of token pricing<br>Cursor receipts currently prioritize tool activity over exact spend accounting<br>Builds a thermal-paper PNG plus a generic share caption<br>Uses skills so the host agent can write the final roast

Kiro CLI session support<br>Kiro CLI is supported as a local session source. Token Receipt reads the Kiro SQLite session store, extracts tool activity and local credit usage, and folds that into the same receipt flow as the other supported agents.<br>Because Kiro does not expose the same local token counters here, the Kiro portion of the bill uses local credit usage instead of a token-derived API estimate.

Experimental Cursor local support<br>Cursor is supported experimentally from local workspace metadata and request trace logs. That gives Token Receipt enough signal to group Composer activity, count reads and searches, and fold those habits into the same receipt flow.<br>The tradeoff is accounting fidelity. Cursor does not currently expose the same local token and cost counters here, so the receipt treats Cursor as behavior-rich but spend-light.

Why this exists<br>The accurate observability product is not the point.<br>The point is a screenshot people instantly understand: your coding agent bill, officially itemized, with a line item for every bad habit you already know you have.<br>Real local signals make the output feel specific instead of generic.

What it itemizes<br>V1 focuses on signals we can defend from the logs:<br>Context window emotional support<br>Repeated file reads<br>Repeated shell confidence loops<br>MCP tool tourism<br>Subagent middle management<br>Planning before touching a file<br>Cache-heavy sessions<br>Low-output expensive runs

Every bill is derived from local agent logs.

How it works<br>Token Receipt is local parsing plus agent-native copywriting, not a new hosted AI layer.<br>1.The runtime scans local Codex, Claude Code, and Kiro CLI session logs, plus experimental Cursor local artifacts

2.Deterministic heuristics turn those logs into receipt facts

3.The skill asks Codex or Claude Code to phrase the final response using the session you are already paying for

No prompt uploads by default. No telemetry in v1.

Who this is for<br>People shipping with Codex, Claude Code, Kiro CLI, or Cursor every day<br>People who want Kiro CLI credits and tool detours in the same receipt

Open source<br>MIT licensed and open source.<br>GitHub repository

FAQ<br>Does Token Receipt need another API key?No for the skill-first path. The local runtime computes the facts, and Codex or Claude Code uses the session you are already in to phrase the roast. The runtime itself does not call OpenAI or Anthropic APIs directly in v1.

Which agents does Token Receipt support first?V1 supports Codex, Claude Code, Kiro CLI, and experimental local Cursor sessions. Cursor support is behavior-first today, so the receipt can reflect local tool activity even when spend fidelity is lower than Codex or Claude Code.

Why is Cursor marked with an asterisk?Cursor support is experimental. Token Receipt reads local Cursor workspace metadata and request traces, so it captures tool activity well, but Cursor does not expose the same local token and cost counters here. That makes the Cursor portion behavior-rich but spend-light compared to Codex and Claude Code.

Does it upload prompts or code?No. The runtime stays local. Skills are instructed to pass sanitized structured facts into the host agent instead of raw session logs unless you explicitly ask for deeper inspection.

Is the receipt supposed to be financially accurate?It is grounded in real local usage and tool signals, but it is still an interpretation of local agent activity rather than a billing ledger. Kiro cost is based on local credit usage rather than token-derived API pricing, and experimental Cursor support currently emphasizes behavioral signals over invoice-grade spend accounting.

What makes this shareable?Each run produces a thermal-paper PNG plus a generic share...