Atlas for Hacker News
Persistent repository context across fresh coding-agent sessions
Atlas indexes your repository locally into a dependency graph, subsystem map, and evidence store — then serves it to Claude Code, Cursor, and Codex over MCP. Built for developers who are tired of re-explaining the same codebase to AI coding agents. Kill the agent, open a new session tomorrow: the context is still there, validated against the current state of your repo.
No signup required · Continue without an account · Local-first indexing · Your code stays on your machine · Windows v1.0
What it does
- Scans a local repository into a resolved import graph, subsystem map, risk model, and symbol evidence.
- Runs Impact, Debug, and Plan Change on the graph, returning the exact cited files behind every result. (In-app repository Q&A is turned off in v1.0.5 — see the changelog.)
- Persists the scan, index, and repository memory to disk, and restores them in fresh MCP sessions after validating a content signature.
- One-click MCP setup for Claude Desktop, Cursor, and Codex — with timestamped config backups that preserve your other MCP servers.
What it does not do
- It is not a coding agent and does not call any LLM. Answers are deterministic lookups against the local index.
- It does not modify your source code. The only files it writes outside its data directory are agent MCP configs, when you click Connect.
- It does not upload repository contents to Atlas servers during indexing — there is no hosted indexing service.
- It does not certify correctness. Impact analysis covers resolved imports; dynamic dispatch and reflection are out of scope.
Persistence, measured
Fresh sessions restore instead of rescanning.
Scans persist under ~/.atlas_desktop/scans/<repo_id>/ (graph, index, risks, evidence, signed memory packet). A fresh MCP process validates a sorted, content-hashed manifest signature against the live repository; if nothing changed, it restores. If anything changed, it refuses the stale scan, tells you why, and asks for a rescan.
Measured on the controlled 27-file benchmark repository that ships in the Atlas source, on one Windows dev machine, using the installed v1.0.1 build: cold scan 2.22 s; restore into a fresh installed MCP process 8–11 ms with no graph rebuild, verified separately for the Claude, Cursor, and Codex launch paths. That is the full scope of this claim — no agent-accuracy or token numbers are claimed.
Data flow, exactly
What leaves your machine.
- Indexing: local process reads local files. Nothing leaves.
- Storage: local disk only (
~/.atlas_desktop). Delete that folder and it is gone. - MCP retrieval: stdio between local processes. Nothing leaves.
- Agent → model provider: the agent you connected may send retrieved context to its provider under that provider's terms. That is your agent's existing data path, not a new Atlas one.
- Analytics: aggregate event names only; never code, prompts, or file paths.
30-second flow
Install → open Atlas → continue without an account → load the sample repository (or scan your own) → ask "Where is authentication implemented?" → get cited files → connect your agent.
Languages
Deep import/dependency analysis: Python and JavaScript/TypeScript. Scanned at file level: Go, Rust, Java, C#, Ruby, and other common source files. Large monorepos index slower — an active work item.
SmartScreen
The installer is not code-signed yet, so Windows SmartScreen warns on first run ("More info → Run anyway"). Verify the download: SHA256 is published below and on the GitHub release.
Founder note
Why this exists
I kept re-explaining the same codebase to coding agents — every new session started from zero, and CLAUDE.md files drifted the moment the code moved. Atlas is the memory layer I wanted: scan once, keep a validated map on disk, let every agent read it, and refuse to serve it when it goes stale. Built solo; feedback — especially critical — is the point of posting here.
Verify the build
Atlas_Setup.exe SHA256:B2531078FC814B9D2AA454FAFD31711AE353AAFCD336C6E39D21B4645A9573EF
GitHub release Changelog Contact
Source-visible project activity: github.com/yoavkozokabab-blip/atlas-repo
FAQ
HN questions
Why not just CLAUDE.md / AGENTS.md?
Hand-written context files drift the moment the code changes, and they hold prose, not structure. Atlas maintains a resolved import graph, subsystem map, and per-file evidence that is validated against the repository signature — if the repo changed, Atlas says so instead of serving stale context.
Why not Cursor's built-in indexing?
Cursor's index serves Cursor. Atlas is one local index that Claude Code, Cursor, and Codex all read over MCP, it persists across sessions with explicit staleness detection, and it adds deterministic impact analysis on the dependency graph — not just retrieval.
Why not Serena or another MCP code server?
Closest neighbor, and if Serena works for you, keep it. Atlas differs in emphasis: persisted validated scans that restore instantly in fresh sessions, deterministic what-breaks impact analysis from the resolved import graph, and a desktop app that writes the MCP config for all three agents in one click.
What does Atlas send to the agent?
Only the context the agent requests via MCP tools: file paths, symbols, graph facts, and short evidence strings. The agent then sends what it chooses to its own model provider under that provider's terms. Full data flow is on the security page.
What is rough today?
Windows-only installer (unsigned, so SmartScreen warns). Deep import analysis is Python and JavaScript/TypeScript; Go, Rust, Java, C#, Ruby are scanned at file level. Pro checkout is not enabled yet.