Architecture
taoCode separates local coding tools from remote model inference. The TAO provider connects those two parts through a consistent gateway interface.
The four layers
Section titled “The four layers”| Layer | Responsibility |
|---|---|
| Agent session | Maintains task context and coordinates model responses and tools |
| Local tools | Read and edit files, execute commands, and return results |
| TAO provider | Resolves a model reference and selects gateway credentials |
| Gateway | Accepts inference requests and serves the selected model |
One provider, many gateways
Section titled “One provider, many gateways”A reference such as tao/chutes/Qwen/Qwen3.8-27B-TEE selects the tao provider, the Chutes gateway, and a model. Changing the gateway changes where inference is sent while preserving the local coding workflow.
Gateways
Section titled “Gateways”A gateway definition describes connection details, authentication, subnet identity, and available capabilities. Model discovery can supplement the curated model list with data from the gateway’s API.
Capabilities should be interpreted at the model and endpoint level. A shared API format does not mean all gateways have the same context limits, billing, or privacy properties.
Extensibility
Section titled “Extensibility”MCP servers can add tools to the agent. Gateway integrations can add inference backends. Both extend the workflow, but they handle different data and need different configuration.
Follow Inference routing for a request walkthrough, or Configuration for a tool-server example.