Neul Labs
High-performance infrastructure for AI agents.
We build the fast parts of the AI stack. Rust-accelerated drop-in replacements for the tools agents actually depend on — LLM gateways, workflow engines, package managers, and developer tooling. Zero config. No rewrites. Just faster.
Projects
brat
Multi-agent harness for AI coding tools. Crash-safe state, parallel execution, one CLI. Orchestrates Claude Code, Aider, OpenCode, Codex, and others through a unified workflow layer backed by an append-only event log.
cargo install --path crates/brat
brat mayor start
brat witness run --once
Rust · Repository
fast-litellm
Rust acceleration for LiteLLM. Drop-in. 3.2× faster connection pooling. 2.8× faster rate limiting. Built with PyO3 and DashMap lock-free data structures. Import it and everything speeds up.
import fast_litellm # that's it import litellm # now Rust-accelerated
Rust Python · Repository · PyPI
fast-langgraph
Rust accelerators for LangGraph. Up to 700× faster checkpoint operations. 10–50× faster state management. Drop-in components for production agent workflows that need to be fast and crash-safe.
Rust Python · Repository
gity
Make large Git repositories feel instant. A lightweight daemon that watches your files via OS-native watchers, implements Git's fsmonitor protocol, and caches status results. git status in milliseconds instead of seconds.
cargo install --path crates/gity
gity register /path/to/large-repo
git status # fast
Rust · Repository
stout
A drop-in replacement for the Homebrew CLI. 10–100× faster for most operations. Pre-computed SQLite index with FTS5 full-text search, parallel bottle downloads, Ed25519 signed updates, and built-in CVE auditing. Works alongside existing brew installations.
stout install jq # instant
stout search json # instant
stout audit # check for CVEs
Rust · Repository
fastworker
Background workers for Python applications. Add task queues in seconds, without the complexity of Celery or RQ.
Python · Repository
Why
The AI agent stack is built on Python libraries that were never designed for production concurrency. LiteLLM's connection pool uses stdlib locks. LangGraph checkpoints serialize through a single thread. Package managers shell out to Ruby.
We fix these bottlenecks one at a time: profile, rewrite the hot path in Rust, ship as a drop-in replacement. Your code doesn't change. It just gets faster.
Contributing
Every project is MIT-licensed and accepts contributions. See the individual repository for setup instructions.
Contact
hello@neul.uk · Built in St Andrews, Scotland.