Skip to content

Platform Capabilities

Heartwood selects one platform adapter from deterministic environment evidence and exposes its capabilities to every interface at GET /project/capabilities. The manifest prevents the terminal, browser, and notebook from advertising routes the platform cannot support.

Current Capability Matrix

Capability Workstation or Container Terra Stanford Carina
Platform identifier generic terra carina
Interfaces Terminal, browser, notebook Terminal, notebook Terminal
Browser route Direct loopback Unavailable Unavailable
Gateway ingress modes Direct loopback (default), Jupyter proxy, trusted proxy Jupyter proxy (default), direct loopback, trusted proxy Direct loopback (default), trusted proxy
Heartwood-managed runtimes llama.cpp, vLLM llama.cpp, vLLM vLLM
Compute model Current host Provisioned Terra compute Slurm allocation
Durable storage Project directory Dedicated directory below /home/jupyter Approved project storage
Credential backends Process, system keyring, mounted file Process, mounted file, managed identity Process, mounted file
Model credential boundary Application-scrubbed Application-scrubbed Application-scrubbed
Model sources Heartwood-managed, Stanford AI API Gateway, ChatGPT sign-in, OpenAI API, Anthropic, other compatible service Heartwood-managed, Stanford AI API Gateway, ChatGPT sign-in, OpenAI API, Anthropic, other compatible service Heartwood-managed, Stanford AI API Gateway
Built-in institution-managed connection Stanford AI API Gateway Stanford AI API Gateway Stanford AI API Gateway

Ingress support describes the gateway contract available to platform operators; it does not make an unavailable researcher interface supported. Terra and Carina continue to advertise only the interfaces in the Interfaces row.

Application-scrubbed routes exclude provider credentials from supported tools and persisted or interface state, but they do not isolate an in-process OpenHands tool from the model client. Secret-backed model routes therefore require Review Every Action on all built-in platforms. Credential-free Heartwood-managed inference does not place a model credential in Heartwood and may use Low-Risk Automation when platform policy permits it.

Startup Projection

GET /project/startup?interface=web and the equivalent gateway method return:

  • selected platform and interface;
  • resolved project and state paths;
  • supported-interface status;
  • current setup phase;
  • next action;
  • browser access URL when one can be verified;
  • compute and confirmation requirements;
  • readiness checks; and
  • the complete platform capability manifest.

Inspection is read-only. POST /project/initialize records explicit browser confirmation and creates private project state.