pacifio/atlas

Atlas – Source‑control‑aware IDE for coding agents

What it is – Atlas is a desktop application (macOS / Windows) that combines a code editor, Git UI, terminal and a knowledge base with a runtime for AI coding agents. Every time an agent (Claude Code, Codex, Atlas’s own native agent, or any agent that follows the Agent‑Client‑Protocol) makes a change, Atlas records a checkpoint that ties the Git commit to the exact prompt, tool calls and reasoning that produced it. The data stay on your machine unless you opt‑in to sync with a team.

Why it matters – Current AI‑assisted development tools lose the “why” behind a change: the prompt, the chain‑of‑thought and any intermediate tool calls disappear once the scrollback buffer clears. Atlas preserves that context, lets you query it later, and lets multiple agents share a common memory so you can switch agents mid‑task without losing progress.

---

Core concepts

Concept	What it does
Checkpoint	When you commit, Atlas links the commit to the agent session that produced it, storing prompts, tool calls, file changes and reasoning in a local SQLite DB (.atlas/sessions.db).
Shared agent memory	An on‑device semantic index (local embeddings + HNSW search) that every agent can read from and write to, enabling one agent to see the plans and decisions of another.
@ mentions	In‑prompt shortcuts that resolve locally to files, folders, symbols, branches, commits, markdown notes, papers or past sessions. The resolved content is injected before the prompt is sent, keeping context windows small.
Knowledge base	Plain‑markdown notes under .atlas/knowledge/ that are versioned alongside the code and automatically indexed for retrieval.
Organizations	Optional sign‑in that syncs projects, checkpoints and knowledge across teammates; everything else works fully offline.

---

Main features (as described in the README)

• Multi‑agent sessions – Run Claude Code, Codex, Atlas’s native agent, or any ACP‑registered agent side‑by‑side in separate tabs; sessions are independent of UI tabs so switching UI never aborts a running agent.
• Agent‑aware Git UI – Full Git graph, staging, branching and diff view; commits are automatically linked to the producing agent session.
• Terminal integration – Each terminal command records its output, exit code and duration; you can run vim, htop, etc. inside Atlas.
• Semantic search & embeddings – On‑device embeddings let you @‑mention relevant knowledge (papers, notes, past sessions) without sending data to the cloud.
• Skills & Packs – SKILL.md files (global or per‑project) and GitHub‑hosted packs can add custom commands, sub‑agents or hooks; they are symlinked into an agent’s skill directory.
• Research tools – Search arXiv / Semantic Scholar, import papers, and @‑mention them directly in prompts.
• Browser tab – Native WebKit view with full login/cookie support and a reader mode.
• Spaces – A JSON‑backed spatial board for visual note‑taking and linking.
• Telemetry – Anonymous usage metrics are on by default but can be disabled; no code or prompts are ever sent.

---

How it works (high‑level flow)

1. User writes a prompt in the chat pane.
2. Atlas resolves any @ mentions (files, notes, past sessions, etc.) locally in Rust.
3. It injects the resolved content, the shared memory snapshot and a short “session handoff” fact‑pack into the prompt.
4. The prompt is sent to the selected agent via the Agent‑Client‑Protocol (ACP) – either as an external subprocess (Claude Code, Codex) or the in‑process native Atlas agent.\nDuring the session, every tool call, file read/write and generated patch is recorded. When the user later commits, Atlas observes the Git operation, matches the resulting patch ID to the stored session and creates a checkpoint entry.

---

Typical workflow

1. Open a project – Atlas creates a .atlas/ folder next to the repo.
2. Start a session – Choose an agent (e.g., Claude Code) and ask it to implement a feature.
3. @‑mention a file or a design doc; the agent receives the path, not the whole file content.
4. Agent makes changes – Atlas applies patches, updates the shared memory index and logs the tool calls.
5. Commit – Atlas links the new commit to the session as a checkpoint.
6. Switch agents – Open a new tab with Codex; the shared memory gives Codex visibility into Claude’s plan, so work continues seamlessly.
7. Later review – Select a checkpoint, open a chat with it, and ask “Why did you change this function?” – the agent replies using the original reasoning.

---

Availability & getting started

• Download – Pre‑built macOS .dmg (Apple Silicon or Intel) and Windows .msi are provided on tryatlas.cc or the GitHub releases page.
• Build from source – Requires Bun, Rust, and platform‑specific build tools (Xcode CLI on macOS, MSVC C++ workload on Windows, GTK 3/WebKit2GTK on Linux). The repo includes scripts (bun run dev:app, bun run build:app:*) for development and production builds.
• Documentation – Full docs at https://docs.tryatlas.cc/ covering chat, memory, editor, Git integration, organization sync, etc.
• Community – Discord server, issue tracker and contribution guidelines are linked from the README.

---

License

Apache License 2.0 (per the LICENSE file).

---

Bottom line

Atlas is a full‑featured, locally‑first IDE that treats AI coding agents as first‑class participants in the software development lifecycle. By persisting prompts, tool calls and reasoning alongside the code, it gives developers an auditable history of how a change was generated and lets multiple agents collaborate without losing context.