James: the AI that routes, checks and improves itself

Never upgrade your model again. New AI models come out every few weeks. With James you don't have to track them, test them or switch. James benchmarks every new model on the kinds of work its users do, and moves it into each role where it proves itself better or cheaper. You keep calling James and simply get better answers. You can still pin a specific model if you want to.

The right model for every request. James picks the best-suited model for each job, often a cheaper open-weights model, and keeps the strongest models for hard work.

It understands the request. James works out what kind of task it is and how hard it is, and sets effort and thinking time to match.

It plans big jobs. Larger tasks are split into parts, and independent parts run at the same time.

It checks its own work. Before answering, James runs the code, tests it, checks answers against the problem's rules, opens apps in a real browser, and has a vision model review the design.

Real tools, not guesses. Web search for facts, code execution for maths, and exact solvers (chess, equations, logic) where a model would only estimate.

Describe an app, get an app. A working, designed, multi-file app with a live preview. James explains what it's doing as it builds, and fixes what it finds broken.

Top-tier quality at a fraction of the cost. Similar quality to the best models at 2 to 10 times lower cost on most benchmarks.

Built to stay up. Automatic failover across models and providers.

Open-weights option. A version that runs entirely on open-weights models.

Works with your tools. An OpenAI-compatible API, so James can be the model in tools like Claude Code, Cline and Kilo.

Fully adjustable. Any setting can be changed per request, including which model leads, effort and timing.

Transparent. Every answer can show how it was routed and why, and builds narrate their decisions.