Checked for the 1.10.0 release candidate on 2026-09-06. An implemented parser is not a promise that every current version of that tool is supported.
Live checked means local numeric usage was inspected during this release work. Fixture tested means the parser is exercised using controlled input. Neither means invoice reconciliation, cross-platform testing, or an exhaustive audit of every historical session.
| Integration | Evidence for this candidate | Scope and limits |
|---|---|---|
| Codex CLI | Live checked + fixtures | New response receipts and legacy cumulative counters; real receipt totals reconciled. Replays and mixed formats covered. |
| Codex Desktop / VS Code | Live checked for response receipts; SQLite fallback fixture tested | New receipts use the granular parser. Opaque SQLite totals are baseline deltas with no invented cost; cannot reconstruct usage before observation. |
| Claude Code | Live numeric reconciliation + fixture coverage in scan/relay tests | Existing day totals preserved during a raw-data rebuild. No paid request or invoice check performed. |
| Cursor | Live numeric reconciliation (2026-10-16, under Node: 76 token-bearing bubbles in state.vscdb summed with sqlite3/jq — input 11,148,181 / output 1,007,118 exact) + fixtures |
Local SQLite cursorDiskKV bubbles; only ~2.5% of bubbles carry token counts. The tokscale CSV path was not present on that machine and was not re-checked. |
| Gemini CLI | Fixture tested | Local token buckets; current live installation not validated here. |
| Qwen | Fixture tested | Local usage only; no provider/model completion performed. |
| Roo Code | Fixture tested | Local usage and reported cost when exposed. |
| Zed | Fixture tested | Public-schema-based reader; not claimed as current live verification. |
| VS Code Copilot | Fixture tested | Activity/model metadata; tokens and cost may not be exposed. |
| Antigravity | Fixture tested | Local activity parsing; opaque data is not converted into invented tokens/cost. Optional live-credit path has separate tests, not live verification here. |
| Muse (Meta) | Live numeric reconciliation (summed against raw JSONL — 203 responses exact) + fixture | Counts only model_completed events and ignores the goal_usage_attribution echo (else double-bills). input_tokens is OpenAI-style (cache-inclusive) — cached portion stripped. Project from workspace_root. |
| Cline | Live data observed + fixture | chat_usage stream, Anthropic-style buckets. The kanban/hub log carries no model or workspace, so model is unknown and project falls back to cline; cost uses Cline’s own figure when present. |
| Augment (Auggie) | Live data observed + fixture | One record per exchange, keep-max across response nodes (placeholder + final pattern). No model id stored → tokens surface, cost pends a model. Project from repository_root. |
| Copilot CLI | Live numeric reconciliation (2026-10-21: per-day totals equal an independent tally of the session.shutdown rollups on all 12 days) + fixtures |
Each assistant.message is an output-only record at its own time; the shutdown adds the input and cache the messages didn’t cover, so a session that runs past midnight puts output on the day it was spent, and one that never shuts down still counts its output. Usage is OpenAI-style: inputTokens includes cache reads (stripped) and outputTokens includes reasoning (kept inside output; before 1.18.0 reasoning was counted a second time). A resumed session’s later shutdown adds only its growth. Distinct from the VS Code Copilot parser. |
| OpenCode | Live numeric reconciliation (2026-10-16: every bucket summed with sqlite3/jq against the raw stores — DB 620 + legacy JSON 305 responses exact, under Bun and Node) + fixtures | Reads both opencode.db (message.data JSON) and legacy storage/message JSON, deduped on message id — the two are disjoint on real installs. tokens.input is already uncached. OpenCode’s own per-response cost is kept as tool-reported; -free/:free models are a reported $0. Before this fix every DB-era turn was dropped. Since 1.18.0 the model id names the provider that served it (deepseek-v4.1-flash@ollama-cloud), and a response with no reported cost is priced at that provider’s rate. |
| Kimi Code (app and CLI) | Live numeric reconciliation (2026-10-21: 909 usage.record calls across main and subagent wires, every day/model total exact against an independent tally) + fixtures |
~/.kimi-code/sessions/{workspace}/{session}/agents/{agent}/wire.jsonl, one record per usage.record (usageScope: "turn"). inputOther is uncached input. Project from the session’s state.json cwd; subagents marked. kimi-code/k3 is priced as kosha’s kimi-k3; plan aliases such as kimi-for-coding don’t name the model behind them, so they count tokens but stay unpriced. |
| Kilo CLI | Fixture tested (shares the OpenCode reader) | Current message.data schema and the older flat messages table. No live installation validated. |
Amp, Droid, OpenClaw, Pi, Kimi CLI (~/.kimi), Kilo, Mux, Synthetic |
Implemented; not individually validated in this candidate | Generic aggregation tests do not establish current parser compatibility. Treat as provisional until a local numeric sample is checked. |
| Custom sources (Ribhu, Chitragupta, Grok, any) | Ribhu live checked (2026-10-18: 158 lines → 158 records, tokens exact per model); Chitragupta live checked (2026-10-18: 2 lines exact, and a Ribhu turn routed through Chitragupta counted once, on Ribhu’s side) + fixtures | Agents with no built-in parser, registered via customSources in ~/.tokmeter/config.json and emitting the canonical tokmeter-usage-jsonl format (one object per API response). Tagged ribhu/chitragupta/grok/custom; model is the resolved model so routed-per-response agents still price. A line’s optional routed-vendor provider marks local-runtime responses (Ollama, LM Studio, llama.cpp, MLX) as a known $0. See the SKILL for the line format. |
SQLite-backed sources and Node
OpenCode, Kilo CLI, Cursor, Zed, Codex Desktop and Antigravity are read from SQLite. Under Bun that is bun:sqlite; under Node (the daemon, the CLI, every npm install) it is Node’s built-in node:sqlite, available without flags from Node 22.13 / 23.4 (22.5+ behind a flag), with better-sqlite3 as a fallback when it is resolvable from the package (not a global install). On older Node these sources report nothing rather than failing — upgrade Node. Before the fix that added node:sqlite, the published package had no Node driver at all, so these sources were empty for every Node user. Cursor was reconciled against raw data under Node on 2026-10-16. Codex Desktop only reports growth on threads updated today that lack JSONL telemetry, baselining a thread on first sight instead of emitting its lifetime total; on that date no thread qualified, so it correctly reported nothing.
Windows
Not supported yet. Since 1.18.0 a CI job builds tokmeter on Windows and runs the CLI and MCP server on Node 20 against an empty home, so a build or startup break shows up. The test suite runs there too, but 51 tests in 10 files still assume a Unix machine (the sqlite3 CLI in fixtures, tar with C: paths, path checks in restore), so it doesn’t gate the job yet. No Windows install of any agent has been checked against real usage.
Keeping this table current
- Capture only structural event fields and numeric usage, with synthetic identifiers and project names.
- Add a regression fixture for any newly observed format. The current Codex fixture is an example.
- Verify totals independently of the parser and test duplicate/replay behavior.
- Record the observation date and what was actually checked. Do not promote fixture-only coverage to live verification because tests passed.
The suite retains pre-existing TODO tests. A passing run is evidence for the tested behavior, not a blanket compatibility certificate.
Something wrong or out of date? This page is generated from its source on GitHub, at the revision this site was checked against.