Track the origin and context of code written by coding agents
When an error occurs in production, you can immediately see which agent received which instructions to write that code. This tool connects session records with generated commits while you use multiple coding agents simultaneously. You can review agent activity on a shared team timeline and intervene in real time if needed.
Checkpoints linking agent sessions and commits
Atlas records agent sessions in a local .atlas/sessions.db file. Secret information is removed before being saved to disk. When a user creates a commit, a checkpoint is generated that links the commit to the session that created it. This link persists through history-modifying operations such as rebasing or amending. Selecting a checkpoint allows you to converse based on what actually happened in that session, restoring context without reading the original transcript.
Shared memory and context transfer
Corrections or decisions injected into one agent are passed to other agents via shared memory. Using the @ symbol to reference files, branches, or past sessions includes that information in the agent's context. Referencing a 5,000-line file transmits only the path rather than pasting the entire content, allowing efficient use of the context window. Embedding search runs on the user's device, so data is not sent externally.
Team collaboration and real-time control
When managing projects at the organization level, sessions sync to the cloud when team members open the same repository. All sessions appear on a single timeline, including agent, model, token usage, and commit link information. Team members can comment on others' sessions or send messages directly to running sessions to adjust agent tasks. You can also ask the organization's agent, Atlas, to summarize team-wide activity or search for specific information.
Local-first design and data ownership
Code, notes, and session records remain on the user's machine by default. External agents communicate with their providers as usual, while only Atlas native agents run through the Atlas gateway. Account creation is optional and should be enabled only when synchronization is desired. Notes are stored in Markdown, canvases in JSON, and sessions in JSONL format, allowing you to close Atlas and work in another editor. Checkpoint records are stored in SQLite files, which are suitable for query optimization.
Pre-use considerations
Atlas runs on macOS 11 or later, Windows 10 or later (x64), and x86-64 Linux with WebKitGTK 4.1 installed. Building from source requires Bun and Rust; macOS requires Xcode Command Line Tools, and Windows requires MSVC build tools. Feature development targets the current version branch, while the main branch is for releases. It is distributed under the Apache License 2.0, and anonymous usage analytics are enabled by default, but code or prompts are not collected.
