Your AI assistant was reading your codebase blind.
Atlas indexes your codebase locally. Now it navigates in milliseconds, not minutes.
No source code leaves your machine. Ever.
The Problem
Your AI assistant is flying blind through your codebase.
A paired benchmark across 50 questions measured lower average tool calls, token use, and latency with Atlas, alongside gains in relationship and impact F1.
33.5% fewer
average tool calls
3.16 baseline → 2.10 with Atlas
26.7% fewer
average total tokens
9,079 baseline → 6,659 with Atlas
50 paired
question-level observations
one repeat per question
The Benchmark
Measured results across 50 paired questions.
JG Business Compliance selected source repositories (medium), tested with gemma4:e4b. These results apply only to the recorded repository, model, adapters, and settings.
| Metric | Baseline | With Atlas | Improvement | 95% paired bootstrap CI |
|---|---|---|---|---|
| Average tool calls | 3.16 | 2.10 | 33.5% lower | [17.7%, 47.1%] |
| Average total tokens | 9,079 | 6,659 | 26.7% lower | [3.1%, 46.2%] |
| Average latency | 8.99 s | 7.78 s | 13.4% lower | [4.5%, 21.8%] |
| Entity accuracy | 83.0% | 88.5% | +5.5 pp | [-0.5 pp, +11.8 pp] |
| Relationship F1 | 54.7% | 87.2% | +32.5 pp | [+9.9 pp, +55.2 pp] |
| Impact F1 | 16.5% | 88.8% | +72.2 pp | [+40.0 pp, +95.4 pp] |
For tool calls, tokens, and latency, improvement is a relative reduction. Accuracy and F1 improvements are Atlas-minus-baseline percentage-point gains. Confidence intervals resample paired question-level means with replacement (seed 44; 5,000 iterations); questions, not tool calls, are the experimental units.
The run used one repeat per question. Entity mentions were extracted from inline-code spans and resolved against the recorded entity catalog; this is not general-purpose named-entity recognition. Relationship and impact scores use explicit, versioned ground truth in the benchmark manifest, not model-generated answers.
Repository: JG Business Compliance selected source repositories (medium). Model: gemma4:e4b. The supplied summary does not identify adapter versions, full model/adapter settings, the per-question list, or exclusions; results should not be generalized beyond this recorded benchmark.
Manifest SHA-256: 6d3032a7cefb866c8dad9bca85567c232c6eaa794920681092a310ec25c1b1e1
Features
Five tools. Complete codebase intelligence.
atlas_ingest
Index a codebase — crawls files, parses ASTs for 18 languages.
atlas_query
Answer natural-language questions about your code.
atlas_impact
Show every function affected by a proposed change.
atlas_edge_check
Detect cross-language bridges and interface mismatches.
atlas_status
Check when a project was last indexed.
18 languages
Python, JS, TS, Go, Rust, Java, C#, C, C++, Ruby, PHP, Swift, Kotlin, Scala, Lua, Elixir, Dart, Zig — via tree-sitter.
Bridge detection
atlas_edge_check detects five cross-language bridge types: HTTP/REST route mismatches, shared environment variable references, config file references, database schema / ORM table bridges, and shared type/interface definitions.
Pricing
One tool call. Zero file reads.
Free
- 1 project
- Python only
- Raw keyword search
- atlas_ingest, atlas_query, atlas_status
- No credit card required
Most popular
Pro
- Unlimited projects
- All 18 languages
- Trained BiLSTM query model
- Full edge-check bridge detection
- Pays for itself in token savings
Installation
One command sets everything up.
One command sets everything up. atlas init handles Meilisearch, indexing, and MCP config automatically.
01
Install
pip install "hornsa-atlas[parsers]"
Python 3.11+ required
02
Run atlas init
atlas init
Downloads Meilisearch, indexes your project, writes MCP config automatically
03
Reload your client
Ctrl+Shift+P → Reload Window
Works with Claude Desktop, VS Code, Cursor, Zed
Stop paying for your AI to read in the dark.
Atlas is free to start. No cloud. No code leaving your machine.
No source code leaves your machine. Ever.