token-pulse

Attribute AI assistant token usage to turns, sessions, commits, branches, PRs, time windows, or files — answering questions like "how many tokens did this PR cost me?"

Reads local Claude Code session logs (~/.claude/projects/**/*.jsonl) for accurate per-turn API token counts (input / output / cache_read / cache_creation), and provides an estimated mode for VS Code GitHub Copilot chat logs.

Status: v0.1 — Claude Code backend is accurate. Copilot backend is best-effort estimation.

Install

# pipx (recommended)
pipx install token-pulse

# or pip
pip install token-pulse

# with optional tiktoken-based estimation for Copilot logs
pip install 'token-pulse[estimate]'

Quick start

# Last conversation turn
token-pulse turn

# Current session
token-pulse session

# A specific commit
token-pulse commit HEAD
token-pulse commit a1b2c3d

# Range of commits
token-pulse commit HEAD~5..HEAD

# Everything since branching off main
token-pulse branch --base main

# A pull request (uses `gh` CLI)
token-pulse pr 1234

# Arbitrary time window
token-pulse window --since "2026-04-20" --until "2026-04-22"

# Per-file attribution from tool calls (Edit/Write/Read targets)
token-pulse file src/foo.py

# JSON output for piping
token-pulse session --json

Sample output

Scope: commit a1b2c3d  "feat: add scoring service"
Window: 2026-04-22 14:31:08 → 2026-04-22 15:08:47 (37m 39s)
Backend: claude-code (12 turns, 1 session)

  Input tokens:           48,210
  Output tokens:          12,883
  Cache read:            312,456
  Cache creation:         18,902
  ─────────────────────────────
  Billable in:            67,112
  Billable out:           12,883

  Estimated cost:         $0.42  (claude-sonnet-4-5)

Why?

Existing tools (ccusage, splitrail, etc.) give you per-day/per-session totals. token-pulse lets you slice the same underlying data by VCS-meaningful boundaries:

Question	Command
How much did this PR cost?	token-pulse pr 42
Which commit was most token-expensive?	token-pulse commit HEAD~10..HEAD --rank
Which file consumes the most agent context?	token-pulse file --top 10
What was the burn rate this morning?	token-pulse window --since 09:00

Backends

claude-code (default, accurate)

Parses JSONL files in ~/.claude/projects/ (override with CLAUDE_PROJECTS_DIR). Token counts come straight from the Anthropic API usage field — exact, including cache hits.

copilot (estimated, opt-in)

Parses VS Code workspace storage chat logs (~/Library/Application Support/Code/User/workspaceStorage/<id>/GitHub.copilot-chat/). Token counts are estimated by re-tokenizing message text with tiktoken (or a chars/4 heuristic if tiktoken is not installed). Does not include hidden system prompts or tool-call payloads, so expect 30–60% under-counting. Useful for relative comparisons, not billing.

token-pulse session --backend copilot

Use as an agent skill

This repo doubles as an agent skill (Claude Code & VS Code Copilot). Drop the repo into your skills dir:

# Claude Code
git clone https://github.com/618coffee/token-pulse ~/.claude/skills/token-pulse

# VS Code Copilot (workspace-scoped)
git clone https://github.com/618coffee/token-pulse .github/skills/token-pulse

Then ask your agent: "How many tokens did the last commit cost?" and it will run token-pulse commit HEAD for you. See SKILL.md.

Configuration

Optional ~/.config/token-pulse/config.toml:

default_backend = "claude-code"
default_currency = "USD"

[pricing.overrides]
"claude-sonnet-4-5" = { input = 3.00, output = 15.00, cache_read = 0.30, cache_creation = 3.75 }  # $/Mtok

Limitations

• Copilot backend is estimation only (see above).
• PR scope requires the gh CLI authenticated to the repo.
• Per-file attribution is heuristic: assistant turn tokens are split equally across files touched by tool_use blocks in that turn.
• Time-window filtering uses entry timestamp; clock skew between machines may cause edge-case misattribution.

Contributing

PRs welcome. See CONTRIBUTING.md.

License

MIT