Your agent rereads the same files all day
Every prompt re-feeds the same modules. Raw reads dump 4,200 tokens when ~920 carry the signal. Tomorrow it reads them again.
Use case · AI coding agents
Stop paying your coding agentto reread your repo.
LeanCTX cuts AI coding-agent token usage by 60–90% by deciding what gets read: AST-aware read modes return signatures instead of full files, cached re-reads cost ~13 tokens, and 95+ shell patterns compress command output. Works with 30+ tools (Cursor, Claude Code, Codex, Copilot) via one lean-ctx setup.
Same engine. Different workflow. Always the right context.
What it costs you today.
Shell output floods the window
One cargo build or npm install can burn thousands of tokens on progress bars and warnings your model never needed.
Context windows fill, accuracy falls
Context-rot research shows model accuracy dropping from 98% to 64% as windows fill with noise. More context is not better context.
The capabilities that do the work.
Everything below ships in the open-source binary today. No roadmap items, no waitlists.
Your tools→LeanCTX→Model
10 read modes
map, signatures, diff, entropy and more. AST-aware via tree-sitter, 26 languages
Session cache
cached re-reads cost ~13 tokens instead of the whole file
95+ shell patterns
cargo, npm, docker, tsc, pytest… compressed to errors + results
30+ AI tools
Cursor, Claude Code, Codex, Copilot, Windsurf, Cline. One setup command
lean-ctx gain
shows exactly what you saved, from your signed local ledger
From zero to first gain.
terminal
# install
$ curl -fsSL https://leanctx.com/install.sh | sh# auto-detect and configure every installed AI tool
$ lean-ctx setup# verify the integration
$ lean-ctx doctor# after a day of work: see what you saved
$ lean-ctx gainOne guide. Two journeys. Full reference.
01
guide · Get started in 10 minutes
Install → setup → doctor → first measured gain.
Learn more02
journey · Daily-use journey
A full working day with LeanCTX in your editor.
Learn more03
journey · Code-intelligence journey
Graph, symbols and AST-powered navigation.
Learn more04
reference · All read modes
Every mode, when it triggers, what it returns.
Learn moreQuestions teams ask before adopting.
Measured on real repo operations: 60–90% fewer tokens per read, ~13 tokens for cached re-reads, and 88–99% on shell output. Run lean-ctx benchmark report . to reproduce the numbers on your own repository.
No. After lean-ctx setup, your AI tool calls LeanCTX automatically via MCP or shell hooks. You keep your editor, your agent and your workflow. The context layer works underneath.
No, and nothing is ever lost. AST-aware modes keep signatures and structure, and every original stays locally retrievable via ctx_retrieve. Smaller context typically improves answers: context-rot research shows accuracy falling as windows fill with noise.
Take back control of your context.
Free for local use, forever. CI enforces it. One binary, ten minutes to the first measured gain.