Adronite launches Codistry AI coding platform, claims half the token cost

Every AI coding assistant has the same dirty secret: it burns through tokens by re-reading huge chunks of your codebase every time you ask it something. Adronite Inc. thinks it has a fix, and it launched that fix today as Codistry, a coding platform built for large enterprise repositories.

The core piece is something Adronite calls the Context Engine, or ACE, which the company has patented. Instead of dumping code into a prompt on every request, ACE builds a relational map of the codebase once, when the software is installed, and updates it as the code changes. When a model needs to work on a task, it only gets the slice of that map relevant to the job. No lengthy prep work on the repository beforehand, according to Adronite.

The company backed the pitch with its own benchmark numbers, which is worth taking with the usual grain of salt since Adronite ran the test itself. Running Claude Opus 4.8 in the cloud through both tools, Codistry used roughly half the tokens Claude Code did on comparable development work, and average cost per task landed about 48% lower. On the open-source project PocketBase specifically, cost per task dropped from $2.12 to $1.10, though Adronite notes that figure folds in a onetime indexing cost spread across later tasks, and both runs used identical prompts, tooling and standard non-batch pricing.

CEO William Colleran, who took over in July after leading RFID chipmaker Impinj Inc., framed the problem as one of context rather than horsepower: AI coding tools stumble not because models are weak but because they don't understand how a codebase actually functions. CTO Edward Rothschild, who co-founded Adronite in Seattle in 2023, put it more bluntly — most competitors solve hard engineering problems by throwing more tokens and bigger models at them. ACE instead hands over only what's needed, leaving the model's reasoning free for the actual problem.

Codistry isn't locked to any one model. It works with frontier models but also lets customers point it at open-weight models they host themselves, and it can run in public or private cloud, on-premises, or fully air-gapped, with source code never leaving infrastructure the customer controls. That's clearly aimed at regulated industries and midmarket firms that can't ship proprietary code to an outside endpoint. To drum up attention, Adronite is also running a 72-hour developer challenge where entrants build an interactive web app on a fixed token budget, judged on efficiency, with a $5,000 top prize and Mac minis for runners-up. The company raised a $5 million Series A in February, led by Gatemore Capital Management.