Atlas Engine · TokenShrinker v1.0.0
Transform massive codebases, versioned docs and web research into high-density, token-optimised payloads for Claude, ChatGPT and local models. Parsing and indexing happen on your machine — your source is never transmitted.
Available for: Windows (.exe) · macOS (.dmg) · Linux (.AppImage / .deb)
Six parts, all running locally. Each figure below is reproducible inside the app on your own prompts.
Tree-sitter compiled to WASM parses your repository into a symbol graph, and Louvain community detection groups it by module. You navigate a map instead of guessing which files matter.
JavaScript · TypeScript · Python · Go · Rust
Source never leaves the machine
Node size is proportional to token footprint, so expensive context is visibly expensive before you send it. Click a subgraph to cut it; a running budget shows the cost as you go.
Presets: signatures only · callers of X · changed since main
Every cut is listed and reversible
Imports are resolved against package.json, requirements.txt, go.mod and Cargo.toml, then linked to docs for the version you actually have installed — not whatever the model remembers.
npm · PyPI · Go modules · crates.io
Private scopes excluded by default
Live search results are stripped of navigation, cookie banners and boilerplate, then run through the same structural compressor as your prompts before they reach a model.
Measured: 34% on JSON, 24% on logs, 35% on traces
Queries sanitised of local identifiers
Indexing is SQLite on your disk. Embeddings run through your own Ollama instance. There is no cloud parsing step to opt out of, because there is no cloud parsing step.
SQLite + sqlite-vec · Ollama embeddings
We could not read your code if we wanted to
Atlas runs a Model Context Protocol server in the background, so Cursor, Claude Code and local agents can query subgraphs directly instead of being handed whole files.
Cursor · Claude Code · custom agents
Runs on 127.0.0.1
A worked example on one task. Drag the slider to see how the payload changes as you cut deeper.
TASK
Add rate limiting to the checkout endpoint
Raw context
18 files80,000
tokens · $0.240
Everything imported by the handler, transitively — plus tests, types and fixtures the model never reads.
Atlas selection
41 symbols4,200
tokens · $0.013
The handler, its middleware chain, the config type, and version-matched docs for the limiter package.
Reduction
19.0×
Tokens cut
94.8%
Per 1,000 calls
$227
How this is calculated: a worked example on one task, priced at $3/M input tokens. This is a retrieval ratio — sending the symbols a task needs instead of every file that touches them — not lossless compression. Cut too deep and the model loses something it needed, which is why every selection shows what was removed and can be re-expanded in one click. Shadow Mode measures the real figure on your own repository before you rely on any of it.
The proxy sits between your tools and the provider. Anything it does not handle passes through untouched, and any single call can opt out with a header.
# Any OpenAI-compatible client
OPENAI_BASE_URL="http://127.0.0.1:1337/v1"
# Anthropic SDKs append /v1 themselves
ANTHROPIC_BASE_URL="http://127.0.0.1:1337"Point it at a folder and watch the graph build. Nothing uploads, and Shadow Mode measures the savings before you change how you work.