Providers and discovery

Connect an existing service after collecting evidence. For narration, PerfChecker supports llama.cpp/Chat Completions, Ollama /api/chat, and an explicitly selected MCP HTTP tool. A Chat Completions URL is not an MCP server. No provider is required for deterministic advice.

Save your actual endpoint/model in perf/advisor.json, then inspect inventory before sending evidence:

{
  "protocol": "chat_completions",
  "endpoint": "http://127.0.0.1:8080/v1/chat/completions",
  "model": "YOUR-INSTALLED-MODEL",
  "timeout": 90,
  "max_evidence_chars": 12000,
  "allow_remote": false,
  "api_key_env": ""
}

using PerfChecker, HTTP
config = load_advisor_config("perf/advisor.json")
setup = advisor_setup(config)
display((setup["status"], setup["evidence_sent"], setup["generation_tested"]))
advice = read_advice("results/advice.json")
narrative = narrate_advice(advice; config) # explicit provider request
display(investigation_view(narrative))

Discovery has evidence_sent=false and generation_tested=false; inspect status and selected availability. Narration does send bounded saved evidence. Read its status, cards and deterministic fallback; generated prose remains unverified.

Actual discovery receipt

Controlled HTTP discovery

HTTP activates provider transport. This complete local discovery fixture returns one model name; it never loads a model or requests generation:

using PerfChecker, HTTP, Sockets
listener = listen(ip"127.0.0.1", 0)
port = getsockname(listener)[2]
close(listener)
server = HTTP.serve!("127.0.0.1", port; verbose=false) do request
    HTTP.Response(200, ["Content-Type"=>"application/json"],
        "{\"data\":[{\"id\":\"controlled-local-fixture\"}]}")
end
try
    config = AdvisorConfig(endpoint="http://127.0.0.1:$port/v1/chat/completions",
        model="controlled-local-fixture", timeout=45)
    result = advisor_setup(config)
    display((result["status"], result["evidence_sent"], result["generation_tested"]))
finally
    close(server)
end

Observed output: ("complete", false, false), with selected_available=true. The receipt records one GET /v1/models to the controlled loopback provider. This qualifies the discovery/transport path, without assessing model reasoning. For a real provider, replace the fixture with a configured service and inspect actual availability. See advice setup.

This controlled HTTP result is not a real provider conversation or evidence of model quality.

Other transports, trust and model management

Use protocol="ollama" with your existing /api/chat endpoint for Ollama; a Chat Completions-compatible llama.cpp service uses the configuration above. model selects the installed model for those protocols. For MCP, the server chooses its model; supply model/limit arguments only when its actual tool schema accepts them. Remote endpoints require allow_remote=true and HTTPS. Store the credential only in the environment variable named by api_key_env.

PerfChecker does not install services or automatically download weights. The editor, Web and Pluto setup controls can discover and save configuration. Local Ollama management supports explicit confirmed download/delete and unload; deletion affects every project sharing that server. Remote management is disabled. Reported sizes include shared layers; cancellation can leave partial server downloads.

An advisor configuration file takes precedence over editor provider settings; a connected temporary local session takes precedence until disconnected. Web studio persists its panel configuration under its reports directory; Pluto action buttons explicitly launch requests. Changing fields launches none.

For another protocol, extend advisor_transport(::Val{:my_provider}, config::AdvisorConfig, request::AbstractDict), translating the bounded request and honoring the timeout. Return the declared response contract; include usage only when the provider reports it.