A Self-Hosted Coding Assistant for Private Repositories

Define what private means for your team

A self-hosted coding assistant can keep model inference local, but privacy depends on the full path: which files enter context, where tool output goes, whether a cloud provider is selected, and how credentials are stored.

Start by deciding which repositories must remain offline and which can use a provider account. A good setup makes that choice visible at the workspace level instead of burying it inside an extension setting.

Keep context narrow and observable

An agent does not need every file in a repository to answer a useful question. Begin with the current file, a diff, a directory listing, or a targeted search. This improves relevance and reduces unnecessary exposure even in a local workflow.

Truss keeps the agent’s workspace tools explicit. You can see tool activity, use read-only modes for investigation, and reserve mutation for a mode and policy you selected.

• Use Chat for a non-mutating codebase question.
• Use Plan for read-only investigation and a proposed approach.
• Review tool activity and Git changes before treating a task as complete.

Separate credentials from repository configuration

Never put a provider key into a committed project file just to make a coding assistant work. Truss keeps provider profiles and credentials in client-side credential storage or environment-based setup, not in workspace source control.

That distinction also makes it safer to share a repository configuration with teammates: the repo can describe the project while each developer chooses a local model or their own provider account.

# Keep provider credentials outside the repository
export OPENROUTER_API_KEY="..."
truss-harness chat --profile cloud

Use a planned fallback, not a surprise one

Some tasks will exceed a local model’s capabilities. That does not mean a self-hosted workflow failed. It means your policy should identify when a task can move to an approved BYOK provider and what context is acceptable to send.

The durable win is control: local for the work that belongs local, an approved cloud profile for the work that needs it, and the same tool and permission model in both cases.