Bring Your Own Model

How it works

All of Driver's model traffic runs through a proxy. Enabling BYOM gives your organization its own identity on that proxy, so your calls carry your credentials, your spend is attributed to your organization, and your budget and rate limits apply to them.

This works on both a shared deployment and a single-tenant private deployment. On a shared deployment only your organization's traffic runs on your credentials; other organizations are unaffected.

What runs on your models

Driver addresses models by role, and every role runs on your credentials once BYOM is configured. The roles fall into four groups:

• Agentic and runtime models. The reasoning models behind Gather Task Context, including a faster tier for latency-sensitive turns.
• Transpiler tiers. The models that analyze and document your codebase during onboarding and ongoing updates, from the highest-capability tier down to the smallest.
• Ticket analysis models. The context-gathering and synthesis models used by Ticket Analysis.
• Embeddings. The embedding model used for search and indexing.

Named credentials

A named credential is one set of provider details that overrides can point at. You can add several and use different credentials for different models. Each one has:

• Name. A label you choose, for example acme-anthropic. Overrides refer to the credential by this name, so it can stay stable while the key behind it rotates.
• Provider. OpenAI, Anthropic, Azure OpenAI, or a LiteLLM-compatible gateway. The gateway option is how you reach a provider Driver does not list directly, or your own internal model routing.
• API key. Stored encrypted. It is never shown back after you save it.
• API base URL. Optional. Set it when your endpoint is not the provider's public one, for example a private or regional endpoint. The endpoint has to be reachable from your Driver deployment, so a private endpoint needs network access such as PrivateLink in place first.
• API version. Optional, and used by Azure OpenAI, for example 2025-04-01-preview.

Model overrides

An override maps one of Driver's model roles to a credential and a target model on that credential, for example pointing Driver's primary reasoning model at anthropic/claude-opus-4-6 on your own Anthropic account.

Rather than mapping each role by hand, route all Anthropic-backed models, or all OpenAI-backed models, to a credential in one step. Individual rows can then be pointed at a different target where a specific role needs one.

Each override can also carry its own requests-per-minute and tokens-per-minute ceiling. Use this when one model on your account has a tighter provider quota than the rest.

Budget and rate limits

These apply across every LLM call your organization makes, on top of any per-override limits:

• Max budget. A spend ceiling in USD for the budget period.
• Budget period. The window the budget resets on, for example 30d.
• Requests per minute and tokens per minute. Throughput ceilings. Leave either blank for no limit.

Setting a budget is worth doing before a first large onboarding run, since that is the most token-intensive thing Driver does against a new codebase.

Cost estimates

Budget tracking needs a price for each model. When you override a model, Driver applies a reference price for the role it replaced, and you can set your own input and output price per million tokens on any override.

Set your own prices when you are pointing at a model Driver would not recognize, such as one exposed by your own gateway, or when your negotiated rate differs from list price. These figures drive Driver's cost estimates and budget enforcement, not your provider invoice, so treat them as approximations rather than billing records.

If your endpoint is unavailable

If Driver cannot resolve your routing, for example a credential is misconfigured or the endpoint cannot be reached, the call falls back to Driver's default model configuration rather than failing. Driver stays usable while a credential problem is sorted out.