stria

A structural codebase indexer for LLM agents. Sub-millisecond queries, zero parsers, any language.

LLM agents routinely guess the wrong file paths because they lack structural context. Existing tools solve this by running hundreds of AST parsers and graph databases. They take minutes to index, consume hundreds of megabytes of RAM, and break on unsupported languages.

Stria takes a radically different approach. It uses grammar-free phrase extraction to index your entire codebase in milliseconds. It compiles to a 4.8MB static binary, uses a bundled SQLite database, requires zero configuration, and works on any language ever invented.

Highlights

• Lazy Context Expansion: Stop blasting thousands of tokens into your LLM's context window. Stria gives agents a compact 50-token structural map of a file and lets them lazy-load full function bodies only when needed via the expand_body MCP tool.
• Blazing Fast on Massive Repos: Stria indexes the entire 3.1GB Linux kernel (72,000 files) from scratch in just 80 seconds. On standard repositories, the build is near-instant (~0.16s) and queries execute in sub-milliseconds.
• Zero Parsers: There are no tree-sitter grammars to maintain. Stria parses C, Rust, Erlang, Nix, or Python exactly the same way, meaning it works out-of-the-box on legacy systems and esoteric languages.

Table of Contents

• Install
• Quickstart
• Agent integration
• Token efficiency
• MCP tools
• CLI usage
• How it works
• Benchmarks
• Limits

Install

cargo install stria

You can also download a pre-built binary for macOS, Linux, or Windows from the releases page.

Quickstart

cd my-project
stria setup --yes
stria serve

The index builds automatically on the first run. There is no init step, no config file, and no per-language setup.

Agent integration

stria setup

This detects installed agents (OpenCode, Claude Code, Cursor, Windsurf) and adds a global Stria entry to their configuration files. You do not need to manually edit JSON.

One switch_repo MCP tool lets the agent change projects mid-session without restarting the server.

When an agent starts a session, it calls orient first. The response includes the language breakdown, a tool guide with use_when hints, and workflow suggestions. The agent can then use:

1. code_search(task="...") to find the right file, its test, and the structural risk. Default tier paths are compact. Agents can add expand_plan for blast radius and coupled files.
2. who_calls(name="...") before refactoring a function to find every file that references it.
3. hidden_deps(file="...") to check whether a change reaches outside the current module.

Each tool returns structured JSON. The agent reads it directly.

Token efficiency

Stria is built to maximize LLM context windows. Graph-based MCP servers often consume 3,000 to 8,000 tokens to answer a single codebase query. Stria compresses structural intelligence into tight, deterministic JSON payloads:

• orient: ~374 tokens
• code_search (default): ~35 tokens
• code_search (expanded): ~75 to 120 tokens
• search: ~100 tokens
• who_calls: ~50 tokens

By separating structure from content, Stria gives the agent exactly what it needs to know, right when it needs to know it, at a fraction of the cost.

MCP tools

CLI usage

stria build --repo <path>     Build or rebuild the phrase index
stria search --repo <path>    Search the index from the terminal
stria serve --repo <path>     Start the MCP server (auto-builds if needed)
stria watch --repo <path>     Watch for file changes and rebuild automatically

How it works

Stria reads all source files as raw text, splits them on delimiter boundaries, counts phrase frequency per file, and applies left-context entropy to classify definitions versus usage.

Searches use IDF-weighted exact match for precision, with BM25 prefix and substring tiers for fuzzy matching. Each file gets optional multipliers for source path, test path, dependency path, and definition density. The index lives in a SQLite database at .stria/phrases.sqlite. Rebuilding is incremental, costing about 0.02s for unchanged files.

Benchmarks

The Linux kernel is the ultimate stress test. It contains 72,000 files spanning C, assembly, Python, shell scripts, makefiles, device trees, and documentation.

Stria indexes the entire 3.1GB repository from scratch in 80 seconds on an Intel 155H. There is no AST parsing, no language configuration, and no memory bloat. It answers complex structural queries across the whole kernel in 170 milliseconds, running entirely from a 4.8MB static binary.

Limits

• Ranking for massive definition files (like the Erlang standard library gen_server.erl at 1,278 phrases) hits a BM25 length normalization ceiling. The correct file is always in the index, but it may rank at 10 or 20 instead of 1.
• Stria is not a linter, bug finder, or security scanner. It measures vocabulary overlap, not code correctness.

License

MIT