Field	Description
Intent	What the session set out to do
Outcome	What actually happened
Intent Score	0.0–1.0 alignment score (via Haiku)
Cost	Token counts and USD estimate by model
Tool Calls	Every tool invocation with input/output
Side Effects	File writes, deletes, git pushes, API calls
Duration	Wall clock time

How it works

Claude Code hooks fire on lifecycle events — no prompting required:

Hook	What happens
`SessionStart`	Creates a session row in local SQLite
`PostToolUse`	Logs each tool call (name, input, response)
`PostToolUseFailure`	Logs failed tool calls with error
`SessionEnd`	Finalizes the receipt: aggregates calls, detects side effects, calculates cost

Receipts are stored locally at ~/.agent-receipt/receipts.db.

Commands

agent-receipt init              # Set up hooks + MCP in current project
agent-receipt init --dry-run    # Preview without writing files
agent-receipt setup             # Install hooks globally (~/.claude/settings.json)
agent-receipt setup --local     # Install hooks in current project only

MCP tools

The MCP server exposes tools for explicit overrides on top of automatic hook tracking:

receipt_start — Set intent, tags, and budget for the current session
receipt_log_tool_call — Log a tool call manually
receipt_end — End session and generate receipt with optional outcome
receipt_score — Submit intent score after evaluating the scoring prompt

Receipt format

Receipts export as Markdown for pasting in PRs:

## Agent Receipt `rec_a1b2c3d4`

**Intent:** Add user authentication
**Outcome:** Executed 12 tool call(s): wrote auth middleware, added tests...
**Intent Score:** 0.85 (via host)

**Cost:** $0.42 · 28,400 tokens · 3m 12s

**Side Effects:**
- File writes: src/auth.ts, src/middleware.ts, tests/auth.test.ts
- Bash: npm install jsonwebtoken
- Git: committed changes