Atlas — local code intelligence for developers and AI agents
Atlas indexes repositories locally and returns focused, source-grounded context about symbols, callers, references, routes, and change impact through CLI, MCP, and HTTP.
GitHub · Download v0.1.46 · Benchmark data (JSON)
Get started
Install with brew install --cask aziron-ai/atlas/atlas or from the public
npm registry with npm install -g @aziron/atlas. The
installation guide also covers the authenticated
GitHub Packages coordinate.
Then run atlas index .,
inspect readiness with atlas status, and preview assistant setup with
atlas bootstrap --dry-run.
Product documentation
- Getting started — first index, query, and assistant connection.
- Installation — Homebrew, npm, archives, and Linux packages.
- Indexing and reindexing — freshness, watch mode, and recovery.
- CLI reference — search, symbols, relationships, impact, and routes.
- AI assistant setup and MCP tools.
- Troubleshooting, privacy, configuration, and upgrades.
- Browse all documentation →
Headline results — Benchmark & Field Comparison — July 2026
- Atlas F1 0.757 at 21.2 context tokens, mean across all 37 languages (native, real-LLM scored, 222 cells / 666 model calls).
- F1 1.000 at 27.1 tokens on the 28 fully-supported languages — full-file-dump accuracy at 6.1× fewer tokens.
- Graph-tool comparison: 0.539 F1 at 97 tokens — Atlas delivers 6.4× the accuracy per token and 36× fewer query tokens.
- Real repository (sirupsen/logrus, gopls call-hierarchy ground truth): Atlas F1 0.975 vs graph tool 0.084 vs raw file 0.017.
- Query latency ~7.4 ms, flat from 15 to 39,161 symbols across 36 real repositories.
- Corroborated by an independent Linux re-run: 37/37 languages fixture-perfect, 9× faster queries than the graph tool, gopls-truth F1 0.933.
Atlas vs Graphify — scorecard
| Metric | Atlas | Graphify | Advantage | Evidence |
|---|---|---|---|---|
| Answer accuracy (F1) | 0.757 | 0.539 | +40% | fixtures · 37 langs |
| Query latency | 7.4 ms | 128 ms | 17× faster | real repos · 36 langs |
| Query tokens (per answer) | 2.6 | 92.8 | 36× fewer | real repos · 36 langs |
| Answer context size | 21.2 tok | 96 tok | 4.5× smaller | fixtures · 37 langs |
| Cold index build | 0.62 s | 1.39 s | 2.3× faster | real repos · 36 langs |
| Incremental re-index | 0.10 s | 1.39 s | 14× faster | real repos · Graphify rebuilds fully |
| Call edges extracted | 9,157 | 1,161 | 7.9× more | real repo · typelevel/cats |
| Coverage vs native AST | 100.2% | 100% | parity | real repo · cats (tree-sitter) |
What a real agent actually spends — Agent-harness token benchmark — 2026-07-10
Claude Code and OpenAI Codex, run headless in sirupsen/logrus (@a23d315) and restricted to one code-intelligence CLI per run; token numbers are each harness's own usage accounting over 19 caller questions with gopls call_hierarchy (LSP-truth), frozen at the pinned commit. Cross-agent absolute totals are not comparable — the two harnesses use different tokenizers and system-prompt floors (see each agent's calibration). Compare modes within an agent.
| Agent | Context source | Mean total tokens | Mean tool calls | Mean F1 |
|---|---|---|---|---|
| claude (claude-sonnet-5) | Atlas | 61,561 | 2 | 0.882 |
| claude (claude-sonnet-5) | No tool (raw exploration) | 144,385 | 4.9 | 0.569 |
| claude (claude-sonnet-5) | Graph tool | 370,898 | 10.3 | 0.305 |
| codex (gpt-5.6-sol) | Atlas | 27,981 | 1 | 0.881 |
| codex (gpt-5.6-sol) | No tool (raw exploration) | 48,106 | 2.1 | 0.876 |
| codex (gpt-5.6-sol) | Graph tool | 109,662 | 9.6 | 0.410 |
Reproducible from any machine: the suite ships in agent-bench/ (pinned commit, frozen gopls question set, isolation flags baked in); per-run records in AGENT_TOKEN_BENCH_PUBLIC.json.
The --detail knob
Retrieval operations floor at high — an agent is never handed a truncated, and therefore wrong, caller list.
| Level | What the agent sees | Tokens | F1 |
|---|---|---|---|
| --detail low | name + location stub | 3 | 0.00 |
| --detail medium | + signatures / explained names | 20 | 0.67 |
| --detail high (default) | full uncapped caller list | 21 | 0.76 |
| --detail xhigh | + stable IDs, docs, scores | 288 | 0.76 |
Language maturity ladder — 40 code languages
Levels reflect validation depth, not just support — an L2 language still indexes and searches; it simply has not reached verified call-graph resolution. Atlas also indexes ~24 content formats (JSON, YAML, HTML, PDF, …) for search.
| Level | Meaning | Languages |
|---|---|---|
| L5 — Reference-validated | call graph cross-checked against an LSP server or SCIP indexer | C, C++, Dart, Fortran, Go, Java, JavaScript, Lua, PHP, Python, Rust, TypeScript, Zig |
| L4 — Real-repo call graph | who-calls resolved and proven on a real repository | Apex, Astro, C#, Elixir, ETS, Groovy, Julia, Kotlin, R, Scala, SQL, Svelte, Swift, Verilog, Vue, Bash (pending real-repo proof), Blade (pending real-repo proof), BYOND (pending real-repo proof), EJS (pending real-repo proof), Objective-C (pending real-repo proof), Pascal (pending real-repo proof), PowerShell (pending real-repo proof), Razor (pending real-repo proof), Ruby (pending real-repo proof) |
| L2 — Real-repo tested | runs on real code; call graph not yet resolved | Delphi, Terraform |
| L1 — Indexed | parsed and symbols extracted | P4 |
Install
Homebrew: brew install --cask aziron-ai/atlas/atlas ·
npm: npm install -g @aziron/atlas ·
Linux: tar.gz / .deb / .rpm / .apk from releases.
Then atlas index . and atlas mcp --transport stdio for agents.
Every number above is reproducible from the committed artifacts under data/. This static digest is replaced by the interactive page when JavaScript is available.
Community
