Questions, answered.

Atlas is local-first repository memory for AI coding agents. It indexes your repo on your machine and returns cited files for questions, debugging, impact analysis, and change planning.

No. The Windows installer is available directly from the Download page.

Repository indexing runs locally. Atlas does not upload your source code to a hosted indexing service.

Only the context you explicitly copy or send to another tool. If you paste Atlas output into Claude Code, Cursor, Codex, or another service, that service receives what you sent.

No. Atlas gives those agents repository memory and cited local context.

Those tools rebuild context per session from what fits in the window. Atlas keeps a persistent local index of the whole repository — dependency graph, subsystem map, impact paths — and serves cited files from it across sessions.

Hand-written context files drift as soon as the code changes, and they store prose rather than structure. Atlas maintains a resolved import graph and per-file evidence, validates it against a content signature of the repository, and refuses to serve it stale.

Cursor's index serves Cursor only. Atlas is one local index shared by Claude Code, Cursor, and Codex over MCP, it survives fresh sessions with explicit staleness detection, and it adds deterministic impact analysis on the dependency graph.

If another MCP server works for you, keep using it. Atlas focuses on persisted validated scans that restore in milliseconds in fresh sessions, deterministic what-breaks analysis, and one-click MCP setup for all three agents from a desktop app.

None. Atlas does not call any LLM — Impact, Debug, and Plan Change are deterministic lookups against the local index, and your coding agent retrieves cited context over MCP. Your agent brings its own model and decides what context to send to it.

The source is publicly visible on GitHub, but it ships under a proprietary license — it is source-visible, not open source.

The core app is free with no account. $19 is planned for capacity features (unlimited repos, sync, snapshots) built on paid infrastructure — priced to sustain a solo-built product without ads or data monetization. Pro checkout is not enabled until Paddle billing is fully verified.

Because shipping one platform well beat shipping three badly. The core is portable Python; macOS and Linux are on the roadmap after the Windows build proves itself.

Code-signing certificates for a new publisher take time and identity verification, which is in progress. Until then, Windows SmartScreen will warn on first run — verify the SHA256 published on the download page and the GitHub release before installing.

Yes. Impact analysis covers statically resolved imports; dynamic dispatch, reflection, string-based imports, and runtime wiring are invisible to it. Every result cites the evidence used so you can verify before acting.

Yes. Connect writes only the atlas entry, keeps every other MCP server untouched, creates a timestamped backup of your config file first, and validates the resulting JSON/TOML before replacing it.

Indexing means reading source files to build the dependency graph and symbol evidence — locally. Atlas never needs write access to your code and never uploads it during indexing.

Repository scanning, the local index, and cited answers work offline. Your AI coding agent still needs its own network access to its model provider.

In a local data directory in your user profile (.atlas_desktop). Nothing is stored on Atlas servers during indexing.

Yes. Delete the .atlas_desktop folder in your user profile, or uninstall Atlas. Removing it deletes scans, settings, and local session data.

No. Atlas reads and analyzes your repository. The only files it writes are agent config files (for example Claude, Cursor, or Codex MCP config) when you explicitly click Connect.

No. The Windows installer runs with the lowest privilege level and installs per-user.

Atlas scans common source files (Python, JavaScript/TypeScript, Go, Rust, Java, C#, Ruby, and more). The deepest dependency and import analysis today is for Python and JavaScript/TypeScript.

Atlas is built for real-world repositories and indexes locally, so speed depends on your machine and repo size. Typical projects index in seconds to a few minutes; very large monorepos take longer.

Atlas detects that the repository changed since the last scan and blocks stale exports until you refresh, so agents never work from an outdated map.

Atlas uses local MCP configuration for Claude Code, Cursor, and Codex where supported by the tool. The app writes the relevant local config and shows exact file paths.

Yes. Manual JSON copy is available in the advanced manual setup path.

Atlas is available today as a Windows installer. macOS and Linux are not available yet.

Do not assume the installer is signed. Windows SmartScreen may show a warning for new or unsigned apps; confirm you are using the official release link before continuing.

Free includes the core local app, local scan, MCP context retrieval, Impact, Debug, Plan Change, and Map.

Pro is planned at $19/month with a 7-day trial: unlimited repositories and indexing, cloud sync, snapshot history, advanced search, priority indexing, New Pro capabilities as they ship, and priority support.

Only when Paddle credentials are configured on the deployment. If checkout is not configured, the Pro button is disabled or shows a not-configured message.

Paddle is the Merchant of Record for paid subscriptions when Pro checkout is enabled. Atlas does not store payment card details.

Team is coming soon. There is no Team subscription billing today.

Yes, anytime from the billing portal. Access continues to the end of the paid period.

Refund requests for billing or service issues are reviewed if submitted within 14 days of the charge. See the Refund Policy for details.

Yes. Signed-in users can delete their account from account settings.

Email yoavkozlovski@gmail.com. Do not paste secrets or full private source code into email.