How we work
Faster. Smaller. Local.
Every design decision in Twira protects three things: the speed of your laptop, the size of your agent's bill, and the privacy of your code.
01
Local by design.
Your code never leaves your machine. There is no Twira cloud to upload to, no "trust us with your source", no telemetry of what you index.
- Indexing · runs entirely on your machine
- Search · all five modes, local
- Detectors · all 40+, local
- Audit chain · local, cryptographically signed, verifiable offline
- LLM calls · through your key, through your provider, your choice, never via us
02
Compiled for speed.
Twira is a single native binary, written in Rust. No runtime to install, no garbage-collection pauses, no startup tax. Every query runs in milliseconds.
- Memory-safe by design · Rust's borrow checker rules out whole classes of bug at compile time, no use-after-free, no buffer overruns, no data races
- Multi-core parallel processing · every core on your laptop, used, indexing, search, and detection all parallelise
- Predictable latency · no garbage-collection pause, no JIT warm-up, no surprise stop-the-world
- Compiled ahead of time · the first query is as fast as the hundredth
03
Engineered for small machines.
Memory footprint measured in tens of megabytes, not gigabytes. Twira runs comfortably on a developer laptop that's already running everything else.
- Embedded local store · optimised for read-heavy analytical workloads
- Lazy index loading · only what your query needs is in memory
- Incremental updates · file changes re-index only the affected slices
- Battery-friendly · no continuous background scan eating CPU
04
Indexed for scale, on SQLite.
Some indexed-context tools store their graph in a single JSON file. That works for a demo. It breaks at monorepo scale: every file change re-parses the whole blob, memory grows with the codebase, queries crawl as the file gets bigger. Twira uses SQLite, the most-deployed database software in the world.
- SQLite as the indexed local store · embedded in every iPhone, Android, browser, aircraft, and most apps you use daily, engineered for read-heavy analytical queries for 25 years
- Proper indexes on every dimension we query · symbol, path, content, dependency, caller, each query hits an index, not a scan
- Incremental updates · a file change touches only the affected slices; no monolithic re-parse
- Vector embeddings for semantic search · your AI agent searches your code and project lore by meaning, not just exact match, embeddings indexed locally next to the rest of the data
- Beyond 200,000 files in one index · measured on monorepo-scale fixtures
- Cross-language structural fingerprinting · 26 languages analysed in the same index, queried the same way
- Five search modes auto-selected · symbol · path · content · regex · semantic
05
Privacy is the architecture.
Most code-intelligence tools require you to upload your source. We made that impossible, by design. The privacy posture isn't a policy we promise to keep. It's the architecture itself.
- No Twira cloud · there's no server for your code to leak from
- No telemetry of indexed content · we never see what you analyse
- LLM provider keys held in the Compliance Proxy · your agent never sees them; rotations are one config update
- Audit chain cryptographically signed and verifiable offline · RFC 3161 timestamps; no Twira servers required to verify
What we don't do
And that's the point.
Each item below is a design choice that other code-intelligence tools have made. Twira made the other choice, every time.
- Require a cloud subscription to start
- Upload your source to a vendor-managed index
- Charge per line of code analysed
- Need ~1 GB of RAM to run usefully
- Pause your editor for seconds while a runtime warms up
- Lock you into one editor or one LLM provider
- Cache embeddings of your code on third-party servers
- Hide pricing behind a "talk to sales" wall for the basics
One install. Premium engineering. No upload.
Install on a laptop, on a CI runner, in an air-gapped environment. Same binary, same speed, same privacy.