split |

MCP server that indexes source files at the function level.

Instead of loading entire files into context, you load one function at a time. Instead of grepping files, you search across 3000+ indexed functions in a single call.

💡 Why

Every time an AI reads a source file, it loads the entire thing — imports, structs, every function — even if you only need one. This wastes context window on code that isn't relevant to the task.

split fixes this by pre-indexing each source file into per-function body files under .split/. The AI loads a function map (cheap), picks what it needs, reads only that (cheap), then edits the original source file with normal tools. The watcher re-splits whenever the source changes.

Source = truth. .split/ = derived cache. One-way sync. Blow it away anytime; it rebuilds from source.

⚡ Token savings

⚙️ How it works

src/config/parser.rs   →   .split/src/config/parser.skel.rs   (structure)
                           .split/src/config/parser/parse.fs
                           .split/src/config/parser/validate.fs
                           .split/src/config/parser/load_file.fs

data/schema.json       →   .split/data/schema.skel.json        (structure)
                           .split/data/schema/_body.fs

• Skeleton = imports, struct definitions, fn signatures with // §ref placeholders
• Body files = one .fs file per function. First line carries the source path + line range, e.g. // §head src/config/parser.rs:42-89 parse — jump straight from body to source line.
• Watcher = one-way: source change → re-split. .fs files are read-only for agents; edit the source instead.

🛠️ Tools

💿 Install

Terminal

claude marketplace add yesitsfebreeze/split
claude plugin install split@yesitsfebreeze

Inside Claude

/plugin marketplace add yesitsfebreeze/split
/plugin install split@yesitsfebreeze

Done. MCP server + skill installed automatically.

🏗️ Building

Requires Rust and the WASM target:

rustup target add wasm32-wasip1
cargo install --git https://github.com/yesitsfebreeze/split

Then wire it up manually. Add to .mcp.json:

{
  "mcpServers": {
    "split": {
      "command": "split",
      "env": {
        "SPLIT_EXT": "rs",
        "SPLIT_SRC_DIR": "src",
        "SPLIT_INDEX_DIR": ".split",
        "SPLIT_MAX_LOC": "256"
      }
    }
  }
}

Bootstrap the index once:

index_dir(src_dir="src", index_dir=".split")

Add to .gitignore:

.split/

Optional: drop a split.ini in the project root instead of env vars — safe to commit.

🔧 Configuration

Place a split.ini in your project root. Safe to commit — no secrets.

SPLIT_EXT         = rs
SPLIT_SRC_DIR     = src
SPLIT_INDEX_DIR   = .split
SPLIT_DEBOUNCE_MS = 500
SPLIT_MAX_LOC     = 256

Priority: env vars > split.ini > hardcoded defaults.

🌐 Languages

split has a WASM language system. Each language is a .wasm module that teaches the parser how to decompose a given file extension.

Language modules live in:

• .split/languages/{ext}.wasm — project-level
• ~/.config/split/languages/{ext}.wasm — user-level
• embedded — built-in (rs, py)

Resolution: project > user > builtin.

Use the list_languages MCP tool to see what is installed in the current environment.

Any language that compiles to wasm32-wasip1 can be a language module. Export:

wasm_alloc(size: i32) -> i32
language_split(ptr: i32, len: i32) -> i32
language_result_ptr() -> i32
language_meta_ptr() -> i32
language_meta_len() -> i32

🧱 Built-in languages

Each language declares its own comment marker and produces a .skel.<ext> skeleton matching the source extension.

Any other extension — whole file stored as one body. Index + search + watch still work; just no fn-level decomposition. Drop a .wasm module into .split/languages/{ext}.wasm to add fn-level support for any language.