TokGuess
Track real AI token usage from local session logs — plus file estimates before you send. 100% local, zero cloud.
How it works
Copilot → VS Code logs (ccreq patterns) → ~75-85% estimated
Copilot (legacy) → ~/.copilot/session-state/*/events.jsonl → real tokens (rare)
Claude → ~/.claude/projects/**/*.jsonl → real tokens
Antigravity CLI → ~/.gemini/antigravity-cli/brain/*/transcript → real or ~estim
Editor → chars ÷ 4.2 → ~90-95% estim
| Layer |
Source |
Accuracy |
Status |
| Copilot Log Watcher |
VS Code GitHub Copilot Chat.log (ccreq) |
~75-85% context-aware estimation |
✅ Active |
| Copilot Watcher |
~/.copilot/session-state/events.jsonl |
Real API tokens |
⚠️ Rare (disabled by default) |
| Claude Watcher |
Claude Desktop *.jsonl (message.usage) |
Real API tokens |
✅ Active if installed |
| Antigravity Watcher |
Google Gemini CLI transcript.jsonl |
Real or content estimate |
✅ Active if installed (Note: Gemini CLI deprecated 18/06/2026, Antigravity CLI is new standard) |
| File Estimator |
Open file in editor |
~90–95% approx. |
✅ Always active |
No cloud. No prompt content stored — only token counts and metadata.
Features
- Context Budget — alert before send:
file + selection = ~8k / 32k tok
- Sidebar dashboard — cost, trends, models, sessions, budget meter
- Status bar — session usage + live budget indicator (🟢🟡🔴)
- Period filters — today / 7d / 30d
- Rescan logs command for full rebuild
Context Budget
Estimates what you'd send to Copilot/Claude before you send:
No selection → full file tokens
Selection → file tokens + selection tokens
| Level |
When |
| 🟢 OK |
Below 75% of threshold (default 32k) |
| 🟡 Warn |
≥ 75% — popup + yellow status bar |
| 🔴 Over |
≥ 100% — warning before context overflow |
Commands
| Command |
Shortcut |
| TokGuess: Open Dashboard |
Ctrl+Shift+T |
| TokGuess: Check Context Budget |
— |
| TokGuess: Rescan Logs |
— |
Configuration
| Setting |
Default |
tokguess.budget.threshold |
32000 |
tokguess.budget.warnPercent |
75 |
tokguess.budget.alerts |
true |
tokguess.sources.copilot.enabled |
true |
tokguess.sources.claude.enabled |
true |
tokguess.sources.antigravity.enabled |
true |
tokguess.sources.copilot.path |
~/.copilot/session-state |
tokguess.sources.claude.path |
~/.claude/projects |
tokguess.sources.antigravity.path |
~/.gemini/antigravity-cli/brain |
tokguess.budget.monthly_cost_usd |
20 |
tokguess.budget.daily_tokens |
50000 |
tokguess.budget.per_file_tokens |
5000 |
Estimation Accuracy
Copilot (Improved June 2026)
TokGuess now uses context-aware estimation for Copilot since events.jsonl is rarely available:
- ✅ Panel/Edit Agent: 4k base input, 10:1 ratio → ~75-85% accuracy
- ✅ Autocomplete: 1k base input, 12:1 ratio → ~80-90% accuracy
- ✅ Chat: 2k base input, 5:1 ratio → ~75-85% accuracy
- ✅ Real model rates: GPT-4o-mini (165 tok/s), Claude Sonnet 4.5 (120 tok/s)
- ✅ Network latency subtracted: 150-300ms depending on model
Previous formula: Fixed 5:1 ratio, generic rates → ~35-45% accuracy ❌
New formula: Context profiles + real benchmarks → ~75-85% accuracy ✅
See AMÉLIORATION-ESTIMATION.md for technical details.
Other Sources
- Claude Desktop: Real API tokens from
message.usage → 100% accurate ✅
- Antigravity CLI: Real tokens from
usage/token_usage fields, or content estimation → 95-100% ✅
- File Estimator: Character count ÷ 4.2 → ~90-95% ✅
Precision Mode (OTel)
Real Copilot token counts via Copilot Chat's official OpenTelemetry export:
- Enable via the 🎯 Precision button in the dashboard — guided 3-step wizard, then Reload Window
- TokGuess sets
github.copilot.chat.otel.enabled + the file exporter: Copilot writes spans (token counts, model, session) to ~/.tokguess/copilot-otel.jsonl, TokGuess tails it — no protobuf, no extra port for Copilot
captureContent stays false: prompts/responses are never exported- A local OTLP collector also runs on
127.0.0.1:4318 for other OTel sources (e.g. Claude Code via OTEL_EXPORTER_OTLP_ENDPOINT)
- Auto-fallback to Chat.log estimation if no spans arrive
- Requires a recent GitHub Copilot Chat extension (otel settings)
Security
The precision collector is hardened by design:
| Measure |
Detail |
| Localhost only |
Both servers bind 127.0.0.1 — no external network, ever. Non-local OTLP endpoints are rejected at startup. |
| Auth token |
Every request to /v1/traces and /api/spans requires a Bearer secret (auto-generated, stored in ~/.tokguess/collector-token, mode 600). Blocks local span injection. /health stays open (no sensitive data). |
| Allowlist scrubbing |
Only known-safe span attributes are kept (token counts, model, session/mode IDs). Prompts, code, messages, and any unknown attribute are dropped before storage. String values capped at 128 chars. |
| Payload cap |
OTLP requests above 1 MB are rejected (413) — token spans are tiny; oversized payloads never get parsed. |
| Strict file permissions |
~/.tokguess is 700, otel-spans.jsonl and the token file are 600 — readable by your user only. |
| No prompt logging |
Never enable OTEL_LOG_USER_PROMPTS — TokGuess does not need or want prompt content. |
Stored data is metadata only: timestamps, token counts, model names, hashed-safe session IDs. No code, no prompts, no secrets.
Disclaimer
- Copilot tokens = context-aware estimation (~75-85%), not official billing
- Claude/Antigravity tokens = real counts from local logs when available
- Context Budget = local file estimate (~90–95%), not official billing
- Cost = public API rates (approx.)
Develop
npm install && npm run compile
# F5 in VS Code
License
MIT
