Atlas

Know what may break
before the code changes.

Atlas maps your repository into persistent, evidence-backed memory — so you and your coding agent (Claude Code, Cursor, Codex) see likely blast radius before a change, locally, across every session.

Windows · Free to use · No signup required · Local-first · Your code stays on your machine

01 · Scan

Atlas maps your repository.

Files, symbols, imports and callers — resolved into a dependency graph and an evidence store. Entirely on your machine.

  • resolving files
  • resolving symbols
  • resolving imports
  • building relationships

02 · Memory

Memory that survives across sessions.

The map is written to disk and validated against the live repo before reuse. A fresh agent session restores it in milliseconds — no re-exploration.

restore 8–11msdependency graph + evidence storevalidated vs live repo

03 · Retrieval

Your agent asks; Atlas answers with citations.

Claude Code, Cursor and Codex request context over MCP. Atlas returns the exact files — cited, ranked, with the reason each was selected. No model in the loop; deterministic lookups.

where is login state handled?
app/_lib/auth.ts0.96session read + cookie sign
app/api/auth/session/route.ts0.91reads the signed cookie
atlas_desktop/static/atlas_accounts.js0.78same account state

04 · Persistent memory

Give your coding agent a persistent
understanding of your project.

Windows · v1.0.5 · MCP · 18 tools · no signup

Capabilities

Read your codebase as a graph.

Three views of the same indexed repository, each grounded in cited files. No model in the retrieval loop — deterministic lookups against the local index.

  1. 01

    Graph

    Where behavior lives, why a module matters, which files are involved — read straight off the dependency graph built from your local index.

  2. 02

    Impact

    The files a change is likely to affect and the dependency paths between them, before you touch a subsystem.

  3. 03

    Debug

    Start from an error or symptom; Atlas parses the traceback to repo files and symbol spans and returns an investigation path.

Integrations

Built for the agents you already use.

All integrations
Claude CodeSupported
CursorSupported
CodexSupported
MCP clientsExperimental
mcp configModel Context Protocol · stdio
{
  "mcpServers": {
    "atlas": { "command": "Atlas.exe", "args": ["--mcp"] }
  }
}

The same local index serves every agent — your source never leaves the machine. Exact per-client setup is in the docs.

Proof

Measured, not marketed.

Methodology & limits
~4sindex · 18-file sample repo
1–30mscited-context retrieval
8–11msrestore across sessions
0.95–0.98file recall · 50 scenarios
18MCP tools

Self-measured on one Windows machine against a reference repository, not a random OSS sample. Every number has a reproducible harness in the repo.

Local-first

Your repo is indexed on your machine.

Atlas scans local files, builds a repository map, and returns cited answers from that local index. It does not upload your repository to a hosted analysis service.

Local Repository index
Local Dependency and impact map
Local Cited answers
You choose Send context to an agent

FAQ

Before you try it.

More questions? Read the full FAQ or contact us.

Do I need an account to download Atlas?

No. Download the Windows installer and start locally — there's a guest mode with no signup.

Does Atlas replace my AI coding agent?

No. Atlas gives Claude Code, Cursor, and Codex a persistent, cited memory of your repository that they can use.

Does Atlas upload my source code?

No. Repository indexing runs on your machine. You decide what context to send to an agent.

How is this different from repeated file search?

Atlas builds a dependency graph + evidence store once and restores it in milliseconds across sessions, returning ranked, cited files — not a fresh grep each time.

Download Atlas.

Windows · v1.0.5 · SHA-256 b2531078fc81… · no signup