██████╗ ██████╗ ██████╗ ███████╗███╗ ███╗██╗ ██╗███╗ ██╗ ██████╗██╗ ██╗
██╔════╝██╔═══██╗██╔══██╗██╔════╝████╗ ████║██║ ██║████╗ ██║██╔════╝██║ ██║
██║ ██║ ██║██║ ██║█████╗ ██╔████╔██║██║ ██║██╔██╗ ██║██║ ███████║
██║ ██║ ██║██║ ██║██╔══╝ ██║╚██╔╝██║██║ ██║██║╚██╗██║██║ ██╔══██║
╚██████╗╚██████╔╝██████╔╝███████╗██║ ╚═╝ ██║╚██████╔╝██║ ╚████║╚██████╗██║ ██║
╚═════╝ ╚═════╝ ╚═════╝ ╚══════╝╚═╝ ╚═╝ ╚═════╝ ╚═╝ ╚═══╝ ╚═════╝╚═╝ ╚═╝
Zero-config, token-efficient code exploration for Claude Code — no MCP server, no Python, no database
The problem
When Claude Code explores a codebase, the naive approach is to read whole files. An 800-line file costs ~8,000 tokens. Read ten files to find one function and you've burned 80,000 tokens before writing a single line of code.
Other solutions add an MCP server, a Python runtime, or a database — all of which consume their own resources and add complexity. codemunch takes a different approach: a pure Claude Code plugin that runs with zero infrastructure overhead.
codemunch fixes this. It auto-indexes your codebase into a flat JSON file and lets Claude fetch the exact 20 lines it needs instead of the 800-line file. No server process, no database daemon, no Python dependency — just install and start using it.
Without codemunch: read auth.ts (800 lines) → ~8,000 tokens
With codemunch: fetch validateToken() → ~35 tokens
Savings: 99.6%
Why use codemunch?
Claude Code bills by token. Every file read costs tokens — and most of those tokens are wasted on code you didn't need to see.
| Scenario | Without codemunch | With codemunch | Savings |
|---|---|---|---|
| Find a function | Read full file (~8,000 tokens) | search + fetch (~47 tokens) | 99.4% |
| Trace all usages of a symbol | Read 14 files (~112,000 tokens) | refs (~45 tokens) | 99.96% |
| Understand project structure | Read 10+ files (~80,000 tokens) | explore (~200 tokens) | 99.75% |
| Typical debugging session (30 lookups) | ~240,000 tokens | ~1,200 tokens | 99.5% |
Benchmarked context savings
We benchmarked codemunch against three real open-source repos — running 5 typical exploration tasks on each and measuring token consumption with vs without codemunch:
| Repo | Size | Traditional | codemunch | Savings |
|---|---|---|---|---|
| expressjs/cors | 6 files, ~8K tokens | 18,810 tokens | 1,281 tokens | 93.2% |
| koajs/koa | 82 files, ~52K tokens | 24,565 tokens | 2,316 tokens | 90.6% |
| fastify/fastify | 290 files, ~468K tokens | 103,891 tokens | 3,174 tokens | 96.9% |
| Combined (15 tasks) | 378 files | 147,266 tokens | 6,771 tokens | 95.4% |
Best case: 99.0% savings (finding all references to .send() across 139 test files — refs instead of grep + reading files).
Worst case: 69.8% savings (multi-fetch of related functions — still cheaper than reading one full file).
Full benchmark methodology and per-task breakdown: docs/benchmark-results.md
That's ~22x less context usage on average, which means:
- Longer conversations — you stay well within Claude's context window instead of hitting limits mid-task
- Lower costs — fewer tokens consumed per task directly reduces your API bill
- Faster responses — less input for Claude to process means quicker replies
- Better results — Claude sees only the relevant code, not thousands of lines of noise, leading to more focused and accurate responses
How it works
codemunch uses a three-tier engine — it picks the best available tool per language and falls back automatically. No configuration required.
Tier 1 — LSP (Language Server)
Byte-exact symbol boundaries, type signatures, cross-file references.
Same engine powering "Go to Definition" in VS Code.
↓ not available?
Tier 2 — Universal ctags
40+ languages. Near-exact boundaries. Ships with most Unix systems.
↓ not available?
Tier 3 — ripgrep
Works on literally any language via regex patterns.
If rg is installed, codemunch works.
Auto-indexing
You never need to manually index. codemunch automatically: