Automated Software Engineering You Can Trust

Translate, verify, and optimize complex software, with independent evidence the output works as intended

why customers use code metal

Stop choosing betweenmoving fast and proving it works.

Legacy modernization
Update, extend, and modernize existing code, with proof that intent is preserved and no new bugs were introduced.
C++ to memory-safe Rust
Code Metal translated and verified 200,000 lines of C++ into memory-safe Rust for a mission-critical customer, preserving the original program’s behavior through formal verification.
57,000 engineering hours saved
8 months faster delivery
Modeling and simulation
Optimize performance
Prototype to production fast

trusted by

One platform, anycode-to-code task

Code Metal automates complete code-to-code engineering workflows, from understanding an existing system to delivering verified software optimized for its target hardware.

Animated Code Metal terminal run view during the translate phase: phases list, work tree of libraries and files building, run reports and log

The Code Metal Approach

Trusted & Efficient

We match models to the work.
We benchmark models on engineering tasks and select for capability, efficiency, and each customer's deployment requirements.
We keep AI within bounds.
We constrain what it can do against explicit requirements, keeping the work focused and its output checkable.
We verify independently.
We use formal methods independent of the model to verify defined properties of proposed changes, giving engineers evidence they can inspect.
We keep engineers in charge.
We automate repeatable work so engineers can shape the system, weigh tradeoffs, and decide what ships.

The Research BehindTrustworthy AI

Methods, benchmarks, and results from our research team

all research

featured

Back in the Race: US Open-Weight Models Closing the Performance Gap

On an internal benchmark translating C++ to Rust, Muse Spark Open reached the top tier of judged translation quality, and strong results across four US-developed model families point to more promising options for teams that need US-developed models to run on their own infrastructure.

read on