An IDE that makes Claude and Codex share their notes

What it does

Atlas is a desktop workspace that coordinates Claude Code, OpenAI Codex, and its own Rust-based agent through a single interface. It runs them in independent tabs, shares context between sessions via an on-device semantic index, and records every prompt, tool call, and file change in a local SQLite store linked to your commit graph. The result is a persistent, searchable history of agent activity that survives rebases and works offline.

The interesting bit

The clever part is the shared memory layer: decisions made by Claude Code automatically surface in Codex’s next prompt because Atlas injects relevant context from an on-device HNSW index before the message ever leaves your machine. Even @ mentions of files, notes, or past sessions are resolved locally in Rust, so the agents get rich context without occupying the context window for the rest of the session.

Key highlights

• Runs Claude Code, Codex, and a native agent in parallel via the Agent Client Protocol (ACP), with sessions that keep streaming if you switch tabs.
• Maintains an on-device semantic memory (local embeddings, HNSW search) that all agents read and write to, plus a knowledge base from markdown notes in .atlas/knowledge/ and existing CLAUDE.md files.
• Records every session to a local .atlas/sessions.db with secrets scrubbed on write, and links commits back to the agent run that produced them—even if you commit from a terminal while Atlas is closed.
• Includes a built-in editor, terminal, git graph with lane assignment, research paper search, and a native WebKit browser, all in a Tauri-based desktop app.
• Local-first by default: works fully offline with no account; optional organisation sync for teams.

Caveats

• macOS is currently the only supported platform; Linux and Windows builds are available but explicitly untested.
• Organisation-wide features such as shared timeline boards and cross-device agent history are on the roadmap, not yet shipped.

Verdict

Developers already using Claude Code or Codex who want a local, persistent audit trail of agent activity—and don’t mind macOS—should try Atlas. If you just need a single agent in a terminal and don’t care about cross-session memory, it’s probably overkill.