yvgude/lean-ctx

██╗     ███████╗ █████╗ ███╗   ██╗     ██████╗████████╗██╗  ██╗
██║     ██╔════╝██╔══██╗████╗  ██║    ██╔════╝╚══██╔══╝╚██╗██╔╝
██║     █████╗  ███████║██╔██╗ ██║    ██║        ██║    ╚███╔╝ 
██║     ██╔══╝  ██╔══██║██║╚██╗██║    ██║        ██║    ██╔██╗ 
███████╗███████╗██║  ██║██║ ╚████║    ╚██████╗   ██║   ██╔╝ ██╗
╚══════╝╚══════╝╚═╝  ╚═╝╚═╝  ╚═══╝     ╚═════╝   ╚═╝   ╚═╝  ╚═╝

Control what your AI can see. LeanCTX — short for Lean Context — is the context engineering layer for AI agents: one local Rust binary that decides what your agents read, compresses what they send to the model, remembers what they learn, guards what they touch — and proves what they save.

Token savings are the receipt. Intelligence is the product. Works with Cursor, Claude Code, Copilot, Windsurf, Codex, Gemini and 30+ other agents — no config needed.

See it in action:

Read + Shell Map-mode reads + compressed CLI output

Gain (live) Tokens + USD savings in real time

Benchmark proof Measure compression by language + mode

_{All GIFs are generated from reproducible VHS tapes in demo/.}

Why developers use LeanCTX

• Longer useful coding sessions — less context waste = more room for actual code reasoning
• Lower API costs — 60–90% fewer tokens on reads and shell output, cached re-reads cost ~13 tokens
• No more "I already showed you this file" — session memory persists across chats
• Works with your existing setup — one lean-ctx setup command, no config changes needed
• Full visibility — see exactly where your context window budget goes
• Model-agnostic & yours — swap OpenAI/Anthropic/Gemini freely; your context and memory stay local and portable, never locked in a vendor's black box

Saves you tokens? Give it a star — it helps others discover LeanCTX.

Why now — own your context

Models are converging on commodity. The durable edge isn't which model you call — it's your context: what your agents read, what they remember, and what you can prove. And the layer that optimizes and owns that context can't come from the vendor that bills per token or keeps your memory in a black box — it has to sit on your side.

That's the shift behind "agent entities" that live in your chat and remember your company (Claude in Slack, ClickUp Brain): a context login, not a model login — you end up renting your own company knowledge back. LeanCTX is the opposite layer. It keeps the moat yours: local-first, portable (.ctxpkg), and model-agnostic — swap OpenAI, Anthropic or Gemini without losing context or cache. Own your context; don't rent it back.

What it does — the four dimensions of context

LeanCTX treats context as a managed resource, not an afterthought. One binary covers the four dimensions that decide how well an AI agent actually performs:

1. Compression — input efficiency

Your AI agent reads files and runs commands. LeanCTX compresses both automatically.

• File reads: 10 read modes (full, map, signatures, diff, lines:N-M, density:X, …) — cached re-reads cost ~13 tokens
• Target density (density:0.4): SDE-style budget compression — keeps the highest-entropy lines until ~40% of the original tokens remain, deterministic
• JIT disclosure: signatures carries line spans and points at lines:N-M for targeted expansion — outline first, bodies on demand
• Shell output: 95+ shell-output patterns compress git, npm, cargo, docker, kubectl, terraform and more (270 passthrough rules)
• Tree-sitter AST: structural understanding for 27 languages — not just text compression
• Reversible by design (CCR): compression never discards content — pruned or truncated payloads move to a content-addressed store with a deterministic handle, so the model can pull the original bytes back on demand via ctx_expand, ctx_retrieve, an in-band marker, or GET /v1/references/{id}. Five recovery paths →

2. Routing — the right fidelity per read

Not every file needs the same depth. LeanCTX sends the signal, not the noise.

• 10 read modes: from full content down to AST signatures and entropy-filtered views
• Adaptive ModePredictor: learns the optimal read mode per file type from past sessions
• IntentEngine: classifies query complexity so simple lookups stay cheap

3. Memory — context that persists

Context doesn't disappear between chats anymore.

• Session memory (CCP): persist task/facts/decisions across chats — structured recovery queries survive compaction
• Knowledge graph: temporal facts with validity windows, episodic + procedural memory
• Property Graph: multi-edge code graph (imports, calls, exports, type_ref) powers impact analysis and search ranking
• Yours, not the vendor's: memory stays local and portable — export it as a .ctxpkg package and move it across machines or models, instead of locking it in a vendor's black box

4. Verification — control what reaches the model

Performance is accuracy, not just speed. You stay in control of the window.

• Context Manager: browser dashboard with real-time token tracking, compression stats, utilization gauge
• Budgets & SLOs: profiles, roles, per-agent budgets, and throttling policies
• Context Proof (ctx_proof, ctx_verify): 4-layer verification engine with CI drift gates

• Web & Research (ctx_url_read): pull a public web page, PDF, or YouTube transcript into context as compressed, citation-backed text — facts/quotes return claims with a confidence score + source URL, relevance-ranked research-compression distils to a token budget, SSRF-guarded (http/https only)
• Graph-Powered Intelligence: hybrid search (BM25 + embeddings + graph proximity via RRF), incremental git-diff updates
• LSP Refactoring (ctx_refactor): language-server-powered rename, references, go-to-definition via rust-analyzer, typescript-language-server, pylsp, gopls
• Multi-Agent (ctx_agent, ctx_handoff): agent handoff with context transfer bundles, diary system, synchronized shared state
• Archive Full-Text Search (ctx_expand search_all): FTS5-powered cross-archive search over all previously archived tool outputs
• PR Context Packs: lean-ctx pack --pr builds a PR-ready context pack (changed files, related tests, impact, artifacts)
• Context Packages: lean-ctx pack create bundles Knowledge + Graph + Session into portable .ctxpkg files with SHA-256 integrity
• Context Time Machine: lean-ctx snapshot create|list|show|verify|restore|publish|import — git-anchored, ed25519-signed snapshots of the layer state (lineage, ledger Φ, ROI, session) on an append-only timeline; replay them in the dashboard, restore to resume a session (and --git to check out the commit), or publish/import a signed snapshot to share it (concept →)
• Observability: lean-ctx gain --live for real-time savings, lean-ctx wrapped for weekly/monthly summaries (gain --svg/--share for a shareable card or self-hostable page), lean-ctx watch for TUI monitoring
• Verified savings: lean-ctx savings is an auditable, per-event ledger (tokenizer transparency, bounce-netting, tamper-evident SHA-256 chain) — local-only, on by default
• HTTP mode: lean-ctx serve for Streamable HTTP MCP + /v1/tools/call (used by the Cookbook + SDK)

Addons — run the ecosystem through one gateway

You don't have to choose between LeanCTX and the other context tools you already like. An addon wraps any MCP server in a tiny lean-ctx-addon.toml manifest; LeanCTX runs it behind its gateway with one command, then treats what it returns like your own reads instead of just proxying it.

lean-ctx addon search memory   # browse the registry by category
lean-ctx addon add headroom    # installs the upstream package + wires the MCP server, on add
lean-ctx addon list            # what's wired into your gateway

• One command to add — addon add <name> installs the upstream package via its native manager (uv, pip, cargo, npm, brew, dotnet) and wires the MCP server into your gateway. No fork, no recompile.
• Folded in, not just proxied — opt-in post-processing runs addon output through the same pipeline as your code: compress to a budget, spill oversized blobs to a ctx_expand handle, index into BM25 / graph / knowledge. Typed adapters route specific tools straight into ctx_expand, ctx_callgraph and ctx_knowledge.
• Untrusted by default — every addon's output is scrubbed for secrets and tagged untrusted before it reaches the model. Always on, not a flag.
• You stay in control — pin a machine or fleet to verified-only, an allowlist, or off entirely via one [addons] policy a single repo can't override.

The registry spans compression (Headroom, Sophon), code intelligence (Repomix, Serena), memory (Mem0, Cognee, Letta) and reasoning (Sequential Thinking). See the addon guide or browse them all.

Where it's going

LeanCTX is growing from a single context layer into a full cognitive context layer for whole teams: version-controlled context strategy, one unified graph, and a governance layer across many agents.

• Context Time Machine → hosted history — the snapshot engine, dashboard replay, restore, and signed file-based share/import have shipped (see above); next is a ctxpkg.com registry for hosted, versioned context history and a side-by-side model-view ｜ git-diff replay. The temporal axis through everything LeanCTX does — it decides, remembers, guards, proves, and replays. (concept →)
• Context as Code — declarative pipelines, profiles, and policies in TOML, versioned like infrastructure
• Unified Context Graph — code, tests, commits, CI runs, and knowledge entries in a single semantic graph
• Agent Harness — roles, budgets, and tool permissions for multi-agent governance
• Context Observability — SLOs on context consumption, anomaly detection, OpenTelemetry / Prometheus export

The full roadmap lives in VISION.md.

How it works (30 se