tokenAtlas

Desktop analytics for AI coding usage: compare models, projects, subscriptions, billing cycles and observed limits from one local history.

Explore the product and illustrated guide at tokenatlas.dev.

This is the initial public source snapshot, under active development. See VERIFICATION.md for the locally checked scope and limitations.

What you can explore

• Usage by model and project, including token counts and cache usage.
• Subscription costs, billing cycles and observed usage limits.
• API-equivalent costs alongside allocated subscription costs.
• Model speed and time to first token, where the source provides verified measurements.

tokenAtlas keeps provider originals in an archive and calculates a local SQLite register incrementally. The dashboard reads prepared results. API-equivalent cost, allocated subscription cost and provider-reported percentages are distinct measures; missing evidence is not treated as zero or an invoice.

Existing integrations

The source includes integrations for Codex, Claude Code, OpenCode, Grok and Gemini / Antigravity. Integration coverage varies by tool and available source data; inclusion does not mean that every metric or quota reader is available.

Gemini / Antigravity and Grok already have import and limit-reading components, with further work planned below. Gemini CLI account discovery does not currently have a connected quota reader. Verified time-to-first-token measurements are currently available for Codex; other sources remain unmeasured for this metric.

Claude Code cloud history requires valid CLI sign-in and a trusted-device credential for the same profile. Failed imports can leave usage and cost comparisons incomplete; the app reports them in its status and cloud import details drawer. See the Claude cloud import guide for access, recovery and measurement limits.

Planned development

We are developing tokenAtlas into a shared place to compare usage across AI tools:

• Refine Gemini / Antigravity and Grok integrations: improve data collection, limit readings and measurement coverage.
• Expand integrations: bring in measurements from additional tools and providers, and extend coverage of their models.
• Extend performance measurements: broaden speed and time-to-first-token coverage where sources provide reliable evidence.

These are development directions, with no committed release dates. Integrations will expand incrementally as provider data becomes available and can be verified. Suggest a source or describe a missing measurement in GitHub Issues.

Project information

• License: Apache-2.0, with attribution in NOTICE
• Data handling: PRIVACY.md
• Private vulnerability reports: SECURITY.md
• Contributions: CONTRIBUTING.md

Requirements

Windows is the currently verified desktop development path. The launcher uses PowerShell, curl and Bun. Building the native parser from source requires stable Rust and the Visual Studio C++ build tools with Windows SDK. _dev.bat can use a published parser only when its source hash and SHA-256 match the release manifest. Availability of that binary must be verified for the public release; source builds are the fallback. macOS and Linux native sources exist, but this preparation does not certify their installation or runtime. Mobile is not a supported target.

Use a current Bun compatible with the committed lockfile. This preparation was checked with the versions recorded in VERIFICATION.md.

Setting up on a new machine

1. On GitHub, choose Code → Download ZIP and extract the complete archive to a writable local directory on Windows.
2. The launcher downloads the matching Windows x64 parser automatically and checks its SHA-256. Rust/MSVC is needed only when building the parser from source.
3. Run _dev.bat. It checks dependencies, prepares the parser, builds the web interface and starts the local server.
4. Open http://127.0.0.1:7331. On a new installation, confirm the archive and source folders in setup before the first import. Configure account history and subscriptions in Settings.

If the port is already occupied, the launcher stops without terminating the application using it. Close that application yourself before starting tokenAtlas.

An existing .env is not required. .env.example documents optional overrides. OneDrive is optional. Keep the database on local storage, outside the source archive. Never import another user's account declarations or database.

Choose the local database folder during setup or in Settings. The installation saves that choice in private tokenatlas.local.json; machine identity and local preferences live in local-state/ beside the selected database. Moving the database through Settings also copies that state and keeps the original as recovery evidence. Existing managed .env selections are migrated automatically; manual environment setup is not required.

Do not expose the local API to a public network. Provider sessions and archives can contain conversation text and credentials; review PRIVACY.md before connecting or sharing data. Back up private data before upgrades.

Local verification

From the source directory, with Bun and Rust/MSVC installed:

bun install --frozen-lockfile
bun run typecheck
bun test ./src ./scripts --timeout 30000
cargo test --locked --manifest-path native/tokenatlas-parse/Cargo.toml
cargo test --locked --manifest-path native/tokenatlas-monitor/Cargo.toml
bun run build:web

Run builds and installation trials in a disposable copy: build:web writes to public/, and starting the server can create or migrate the selected database. Tests that explicitly require private integration data do not establish a clean installation pass when that data is absent. Verification is local; this project does not use GitHub Actions.

The 30-second default test budget accommodates integration cases that migrate synthetic databases and start native processes on Windows. Explicit performance assertions and individual test budgets still apply.

License, attribution and support

Project-owned code, including the native parser and monitor, is offered under Apache-2.0. Third-party dependencies and fonts retain their licenses; see THIRD_PARTY_NOTICES.md and licenses/. The project name and logo follow TRADEMARKS.md; the code license does not imply endorsement of a fork. Contributions retain their authorship.

The program is provided without warranty under its license. Calculated figures are analytical observations and estimates, not provider billing statements. Support is voluntary: SUPPORT.md. Report ordinary bugs with a small, synthetic reproduction and include the source version and checks you ran.