The missing record between agent sessions
Atlas starts from a problem that becomes obvious only after coding agents become routine. Claude Code may explain an architectural choice in one session, Codex may edit the affected files in another, and a third task may commit the result later. Git preserves the code but not the whole chain of reasoning. Individual agent transcripts preserve fragments, but each tool sees only its own history. Atlas tries to join those fragments into a project record.
The desktop app runs Claude Code and Codex as external subprocesses through the Agent Client Protocol, or ACP, alongside Atlas's own agent. It records prompts, responses, tool calls, touched files, patches, plans, decisions, and failures. A local semantic index selects relevant memories for later prompts. On the first message of a new session, Atlas can also provide a compact handoff from the previous one, even when a different agent produced it.
This is more useful than simply putting several chat tabs in one window. Atlas links sessions to Git commits as Checkpoints, including commits made outside the app. It reconciles those links through amendments and rebases using patch identities, and it leaves genuinely ambiguous squashes orphaned instead of assigning them to the wrong session. That restraint makes the audit trail more credible.
A workspace, not an editor plug-in
Atlas includes a CodeMirror editor, a real Git graph, staging and commit controls, terminals, a Markdown knowledge base, an embedded browser, research search, split views, and spatial note boards. Files, folders, symbols, commits, branches, notes, papers, and prior sessions can be inserted into prompts with @ references. Folder references resolve to paths rather than pasting thousands of lines into context.
The knowledge system is deliberately portable. Notes are Markdown, boards are JSON, sessions can be represented as JSONL, and the editor works on ordinary files. The main exception is Atlas's SQLite checkpoint and session record, which is optimized for queries and stored in the gitignored .atlas directory. Developers can close the app and continue in another editor without converting their project.
Local mode is a meaningful strength. The README says code, notes, sessions, and embeddings stay on the machine, with secrets scrubbed before persistence. An Atlas account is optional until a user wants organization and device sync. Anonymous product analytics are enabled by default, however, so privacy-sensitive teams should read the telemetry document and disable collection as part of setup.
The quick start has a narrow platform target
For the intended user, installation is straightforward: download the macOS DMG from the releases page. Claude Code must already be installed and discoverable on PATH to use that agent. Codex also runs externally over ACP. Atlas's native agent avoids that extra CLI requirement, and existing agent subscriptions continue to provide the model access.
Source builds are a different proposition. They require Bun, stable Rust, Xcode command-line tools, and a first Rust compilation that the README says takes a few minutes. Linux builds add GTK 3, WebKit2GTK 4.1, and GLib headers. Windows and Linux come from the same Tauri codebase but are explicitly untested, so the presence of dependency instructions should not be mistaken for support. An open Windows build pull request documents platform-specific compile failures, reinforcing that distinction.
The app's breadth also creates operational surface. It observes repositories, imports multiple transcript formats, maintains a semantic index, runs child processes, hosts native webviews, and tracks Git history rewrites. That is far more state than a chat sidebar owns. Back up important work normally and verify session capture health rather than assuming every event was recorded.
Rough edges are concrete, not theoretical
The most serious open report concerns background output. Issue #94 says an assistant message arriving after the parent turn is finalized can be silently dropped from Atlas's canonical log and interface, even though the original agent transcript still contains it. Developers who launch detached tests or delegated jobs specifically to report later should treat that as a workflow blocker until resolved.
Other gaps reveal the alpha stage. Open issue #33 says notebook files are unsupported in the file viewer, which makes Atlas a poor fit for common data-science repositories. Issue #80 describes MCP support as uneven: the native agent can read an application-level server configuration, but changes need a restart and the current bridge is centered on stdio. The broader cross-agent MCP experience is roadmap work, not a finished shared configuration layer.
Some problems are being fixed quickly. A first-run issue that could trap non-Claude users behind a disabled input was closed in August, and a timeline report involving impossible session durations was also closed. Rapid fixes are encouraging, but they also show how much core interaction and accounting logic is still settling.
Health and the decision
The repository was pushed on August 14, 2026, three days after the alpha-0.2.6 release. GitHub listed 19 open issues and five open pull requests, with recent work covering permission modes, Windows compilation, UI defects, dependencies, and onboarding. That is active maintenance and active product formation at the same time. The MIT license and detailed architecture, security, telemetry, and contribution documents make evaluation easier.
Atlas is worth trying when the cost of fragmented agent history has become larger than the cost of adopting another workspace. Its checkpoint model and cross-agent memory have a clearer purpose than generic promises about better AI coding. It is not yet the conservative choice for platform-diverse teams or workflows where every background response must be captured. On a Mac, for a developer already switching between Claude Code and Codex, the trade can make sense today, provided the local database is treated as a helpful record rather than an infallible source of truth.