A twelve-person agency runs the agent on local models, because its clients' code cannot leave the building

Meridian is an agency: twelve engineers, nine client codebases at any time, and contracts that forbid client source from reaching a third party. Hosted coding agents were out before the conversation started.

Where they started

Each engineer had a private setup — a local model here, a script there — and none of it was shared. Onboarding a new client meant rebuilding the setup from memory.

What they changed

rovecode's provider registry treats a local runtime as a first-class provider. Meridian points every workstation at two models on a machine in the office, with a hosted provider as the last fallback for the one client whose contract allows it. The policy file lives in each client repository under .rovecode/, so the rules travel with the code and the trust gate asks before a project's own policy is honoured.

They use the Teams plan for the shared knowledge and policy and nothing else: the models are theirs, so credits are barely touched.

What they measured

No client source has left the office; the usage panel shows where every token went. About forty percent of routine changes — dependency bumps, test scaffolding, the migration boilerplate every client needs — are now handed to the agent and reviewed rather than written.

What was hard

Model quality. The local models are a generation behind the hosted ones, and the agent's plan panel shows it: more steps, more corrections. Meridian's answer was to give the agent smaller tasks and keep the judgement calls for people — which, Inés points out, is how they wanted to work anyway.