LiteLLM → MC-1

What changes

Keep the existing client behavior stable first. Introduce the MC-1 base URL, logical model, and server-side credential, then add control-plane constraints only after compatibility passes.

Feature mapping

• Inventory proxy aliases, fallbacks, budgets, callbacks, and custom headers.
• Point compatible clients to MC-1 and map aliases to the logical or eligible explicit model.
• Move governance gradually; do not duplicate retry loops in both layers.

Migration sequence

• Inventory endpoints, models, streaming, tools, retries, and error handling.
• Create a scoped MC-1 project and test key.
• Run an owned non-production fixture through both paths.
• Compare outputs, tool continuation, latency, usage, and failures.
• Canary a reversible workload before broader promotion.

Compatibility example

export MC1_API_KEY="..."
export MC1_BASE_URL="https://api.colomboai.com/v1"
# Keep the client OpenAI-compatible and use model colomboai/mc-1.

Source-specific caveats

• LiteLLM configuration is not imported automatically.
• Choose one authoritative layer for retries, budget enforcement, and provider credentials.

Rollback

• Retain the previous route until acceptance passes.
• Do not translate unsupported provider-specific features silently.
• Record the exact client and integration versions used for validation.