Cercano

Models and providers

What it is

Cercano can connect to OpenAI-compatible Chat Completions endpoints using a configured base URL, credentials, and model. That includes compatible hosted open-weight services and self-hosted inference servers, not just OpenAI itself. Profiles connect these endpoints to Cercano's routing system.

Frontier subscriptions and open models together

Cercano also has provider-specific subscription sign-in paths for ChatGPT and Claude, separate from API-key access to compatible endpoints. You can use a frontier provider for the main conversation and route suitable delegated work to local or hosted open-weight models—or use those models on their own.

Subscription sign-in is not a universal replacement for API credentials. Account eligibility, usage limits, provider terms, and third-party-client support apply; availability can change. An implemented sign-in path is not a promise of provider endorsement or unlimited access. Check the options presented in /config → Cloud and the provider-specific controls available in your installed version.

Why it matters

Choose inference based on capability, price, and where you want work to run—not on whether your coding agent has a bespoke integration with a particular vendor. Change providers or try new open-weight models while keeping the same terminal, sessions, tools, and delegation workflow. Combining frontier and open models lets you reserve your frontier allowance or API budget for work that benefits from it instead of using the same inference route for every task.

Try it

Open /config → Cloud. Configure a Chat Completions profile with your provider's base URL, credentials, and a tool-capable model. Assign the profile to the desired cloud destination in Routing, save the settings, and test a small read-only task. /cloud lists profiles; /cloud use <name> selects the active profile and /cloud key <name> <api-key> updates a key. It is not the profile editor.

Route work deliberately

Cercano separates three choices: what kind of task is being performed, how much model capability it needs, and where it runs. Task classes such as reconnaissance, investigation, mechanical development, and implementation have saved quality and destination settings.

Primary, Secondary, and Local are independent destinations. Each cloud destination has its own model profile and optional backups. Redirects change initial placement; backups respond to failures. Check the reported final route instead of inferring destination, privacy, or cost from a task's default.

Open /config, select Routing, and compare the assignments for Reconnaissance and Implementation. Adjust destination and quality independently, save the changes, then run a delegated read-only task and inspect its reported route.

Run local models

Cercano integrates a managed llama-server runtime for open-weight models. Ollama and an optional managed mistral.rs sidecar are also supported. Runtime setup, model selection, and process lifecycle sit inside the agent experience rather than requiring a separate inference service assembled from scratch.

Open /config and inspect Runtime and Local Models. Choose a supported runtime and a model appropriate for your hardware, allow any required download to complete, then assign suitable work to Local in Routing.

Local execution still has model downloads, disk, memory, compute, and quality tradeoffs. It avoids hosted per-token charges for the inference that actually runs locally; it does not make hardware, electricity, tools, downloads, or other configured routes free.

Controls and limitations

  • Compatibility is not universal certification. Providers differ in tool calling, streaming, context limits, image support, and usage reporting.
  • Image input for compatible Chat Completions/Responses models requires confirmed model vision capability; a compatible wire format alone is not enough.
  • A model that handles plain chat may still be unsuitable for an agent's tool loop.
  • Hosted endpoints receive the content routed to them and have their own pricing and data policies. An open-weight model does not imply local inference.
  • OpenAI Responses and subscription sign-in are distinct provider paths; support for Chat Completions does not imply identical behavior across those paths.
  • Quality selects a model within a destination. It does not grant tools or override placement policy.
  • Secondary is a separate destination, not merely Primary's implicit backup.

See delegation and subagents, token metrics, and the complete feature guide.

Cercano + Cercloud documentation Local review draft