Select, verify, and observe a model without working on PtcRunner's provider implementation.
ptc-host.json installs a model under a stable alias, and
ptc.json selects that alias for its trusted workflow. Generated mission code
does not receive the model route.
What must the model support?
An agent loop needs an endpoint that accepts tool-bearing requests. A model may accept ordinary completions while rejecting the agent's tool contract; PtcRunner reports that separately from a missing model or credential failure.
How do I choose one?
Start with the model in the shipped examples: OpenRouter's
deepseek/deepseek-v4-flash.
Model catalogs and routing change over time, so PtcRunner does not keep a static
compatibility list. Choose a model advertised for tool use, install it under an
alias, and confirm the exact route with ptc doctor PROJECT.json --connect.
Two aliases share the public llm-request call budget. config.max_calls
additionally caps an alias only when it is stricter than that shared budget.
Where do credentials belong?
Bind credentials outside the application, preferably through a named environment
variable or another credential source in ptc-host.json. Then inspect the public
installation without revealing its credential or private endpoint:
ptc models ptc-project.json
ptc doctor ptc-project.json
models names the selector each LLM alias configured; an endpoint-bearing
openai-compat: selector is withheld instead, because it carries the
private address from ptc-host.json.
How do I check connectivity?
Plain doctor is inert. Use the active connectivity probe only when a remote
request is intended:
ptc doctor ptc-project.json --connect
What does a run record?
After a run, the trace records how many model calls were made, which alias they used, reported token usage and cost, timing, and a safe failure class. Prompts and responses are private and appear only when private inspection is enabled.
Start with Install models and tools for one complete workflow. The model and host reference owns selector forms, credentials, cache policy, request parameters, ceilings, diagnostics, and connectivity behavior. See Customize an agent for the model-neutral loop.