Rewriting a whole app in Rust costs months, minute-long compiles and gigabytes of target/. Most of the runtime sits in a few functions anyway.
fast-anything finds those functions, has your coding agent port only them to Rust, proves the port gives the same results on real recorded calls, and swaps it in. One command, fully reversible.
Install
npm i -g @el4cteo/fast-anything # or: cargo install fast-anything
Ports are compiled with your Rust toolchain (rustup).
Use
fast-anything go --agent claude -- python app.py # or: -- node app.ts / ruby app.rb / java App.java / mvn exec:java / gradle run / go run . / php app.php / clang -O2 main.c
Or step by step:
--agent takes claude, codex, or any command that reads a prompt on stdin. Without it, hand TASK.md to whatever you use.
Proof, not vibes
That port looked right but compared bytes instead of characters. 148 of 500 real calls caught it. On top of the recorded calls, fast-anything derives extra cases and lets the original answer them: edge cases (NaN, ±infinity, -0.0, zero and ±1, non-ASCII text, empty and up to 16× bigger inputs) and shrunk inputs (tiny and random slices). Nothing is enabled until every case matches: same values (floats within 1e-9), errors where the original errors. A port that is right but slower per call than the original (calling into Rust costs more than the work saved) isn't enabled either. FAST_ANYTHING_DISABLE=1 runs the original any time.
Results
Real runs on an i5-10500, Windows 11 (C and C++: a 4-core GitHub Actions Linux runner). Reproduce them from examples/.
fast-anything ports the outermost loop it can, not the tiny function called inside it, so the app crosses into Rust once instead of millions of times. Porting only the inner levenshtein gave 35× and 1.9×.
Why not rewrite it all?
- Compile time: each port is a one-function crate. Seconds, not minutes.
- Cache: every port shares one
.fast/target. Delete it whenever you like. - Risk: your code stays as it is. Rust is opt-in per function, checked against real behavior.
Agents
- Claude Code plugin:
/plugin marketplace add EL4CTEO/fast-anything, then/plugin install fast-anything@fast-anythingand/fast-anything:fast python app.py - MCP server (Claude Code, Codex, Cursor…), built into the binary:
claude mcp add fast-anything -- fast-anything mcp. 8 tools:fast_scan,fast_port,fast_verify,fast_enable,fast_disable,fast_bench,fast_report,fast_status. - Skill:
skills/fast-anythingteaches any agent the loop.
Languages
It ports top-level, deterministic functions that take plain data (numbers, strings, lists, dicts, typed arrays…), including ones in your entry script when its main code sits under if __name__ == "__main__": (or if __FILE__ == $0, or if (import.meta.main) / if (require.main === module)). In Python that means module-level functions and @staticmethods, in Java, C# and PHP static methods (and PHP functions), in Go plain functions, in C and C++ free functions and static methods. scan says why anything else can't move yet.
With Intel oneAPI on Windows, run fast-anything from a shell where setvars.bat ran (it sets up the linker). SYCL kernels run on the device and aren't ported; with -fsycl, only host functions are profiled and ported.
Ruby, Java, Go, PHP, C, C++ and C# need no headers, bindings or extra tools: the port is a plain Rust library with one C entry point taking and returning JSON, and fast-anything replays the recorded calls against it directly. Any language that can call C can use that bridge.
License
MIT