slim

CodeCondense — Stop Burning Money on Claude Code

The token-saving plugin every Claude Code developer needs. Cut your API costs 40–83%. Automatically. Locally. Free to start.

npx @codecondense/install

---

Claude Code is powerful. It's also expensive.

Every time Claude reads a file, searches your codebase, or edits code, it burns tokens. Most of those tokens are wasted — duplicate reads, bloated search results, redundant roundtrips that Claude makes because it has no smarter option.

CodeCondense cuts that waste by 40–83%. Without changing how you work.

Without CodeCondense:   7 roundtrips · ~31,500 tokens · ~$1.34 on Opus
With CodeCondense:      2 roundtrips · ~8,000 tokens  · ~$0.34 on Opus
                                                         Same result.

A typical heavy session makes 200–400 tool calls. CodeCondense collapses 40–83% of those.
At Opus rates that's $3–$7 saved per session. Every session.

---

Benchmark Results

Tested on Claude Sonnet 4 (v0.8.36, June 2026) · Pricing: $3/M input · $15/M output

Scenario 1 — Build from Scratch (10 apps, Node.js + Python)

Metric	Result
Fewer tool calls	−85%
Fewer tokens	−72%
Cost saved	$1.44
Time saved	15m 46s

Scenario 2 — Real Codebase (7,256 files · 1,326 .tsx · 1,699 .ts · 192 deps)

Task	Vanilla calls	Slim calls	Vanilla tokens	Slim tokens	Vanilla cost	Slim cost	Time saved
T1 — Simple (2 files)	8	3	38K	11K	$1.20	$0.22	−66%
T2 — Medium (2,478-line file)	7	2	95K	16K	$2.80	$0.44	−83%
T3 — Complex (5 files)	14	3	145K	22K	$4.60	$0.76	−83%
Combined	29	8	278K	49K	$8.60	$1.42	−80%

Combined reduction: −72% calls · −82% tokens · −83% cost · 13m 12s returned

Monthly Projection (8h/day × 20 days, 15 tasks/day)

Metric	Without Slim	With Slim	Savings
API cost	$756	$125	$631 saved (−83%)
Tool calls	2,640	780	1,860 fewer (−70%)
Tokens	24.7M	4.6M	~20M saved (−82%)
Wait time	24.3h	5.1h	19h returned (−79%)

Full methodology and raw data: codecondense.com/benchmarks

---

Features

🌳 AST Symbol Index — real code understanding, not grep

CodeCondense uses tree-sitter — the same parser behind VS Code and GitHub Copilot — to build a full Abstract Syntax Tree of your codebase. It understands what your code means, not just what it contains.

Without CodeCondense:
  Claude → grep "render"   → 847 matches, 203 files, 12 reads, ~54,000 tokens

With CodeCondense:
  Claude → Sql({ op: "search", name: "render", type: "function" })
         → 3 results: exact definition locations, 1 roundtrip, ~200 tokens

270× token reduction. Same answer.

Supports 37 languages: TypeScript · JavaScript · Python · Go · Rust · Java · Kotlin · C# · Swift · PHP · Ruby · Scala · Elixir · C/C++ · Dart · Vue · CSS · Bash · YAML · JSON · TOML · Solidity · Zig · and more — all via precompiled WebAssembly grammars. No setup, no language servers.

---

🔍 Smart Search — grep + read + AST in one call

Without CodeCondense:   Glob → Read → Read → Grep → Read → Read → Read
                        7 calls · ~12,000 tokens

With CodeCondense:      Search({ content_regex: "...", output_mode: "file_paths_with_content" })
                        1 call · ~2,400 tokens

• Ripgrep under the hood — sub-second on codebases with 100k files
• Delta mode — re-reading a file? Returns a unified diff of what changed, not the whole file again (80–95% token savings on repeated reads)
• AST summary mode — understand a 600-line file in 40 tokens
• BM25 ranking — results sorted by relevance density
• Budget enforcement — results capped at a token budget, never floods context

---

✏️ Batch Edit — all your edits in one roundtrip

Without CodeCondense:   5 separate Edit calls · 5 waits · ~15 seconds · $0.20 on Opus

With CodeCondense:      Edit({ edits: [file A, file B, file C, file D, file E] })
                        1 call · 1 wait · ~3 seconds · $0.04 on Opus

Every edit is guarded: exact-match validation, pre-write backups, shrink fuse, empty-write fuse, symlink guard, Jupyter notebook support.

---

🔬 Investigate — the search-everything command

Claude → Investigate({ query: "useAuthContext" })
       → symbols:  exact AST definition (file:line, type)
       → matches:  all usages across the codebase
       → files:    top files ranked by density, each with AST symbol map

Collapses the entire explore-a-symbol workflow into a single tool call.

---

🧠 Session Memory — Claude that actually remembers

Claude Code starts every session with amnesia. CodeCondense fixes that.