ModelsThe model server

Address and rules

• Where: 127.0.0.1:36434, which RoachNet.app always uses. If another program already holds 36434 (usually an ollama serve that an older RoachNet left running), RoachInference listens on a free loopback port instead, and GET /api/system/ai/providers names it. A runtime you start yourself listens wherever OLLAMA_BASE_URL points, as long as it's a loopback address: Ollama's port, 11434, or no port means 36434. It won't start the model server anywhere else.
• Auth: no token, like Ollama. Instead, the Host header must be a loopback name, and an Origin header must be a loopback, app://, file://, tauri:// or vscode-webview:// origin. Other websites get a 403.
• Bodies: may be sent without a JSON content type, as Ollama clients do.
• Errors: OpenAI's shape on the OpenAI routes, Anthropic's on /v1/messages, Ollama's everywhere else. A missing model is a 404 before any stream starts.

MODELS=http://127.0.0.1:36434
curl -s "$MODELS/v1/models"

Ollama routes

Chat and generate stream newline-delimited JSON unless you send "stream": false. Tool calls come back as objects and are never streamed half-written. format takes "json" or a JSON Schema, which is enforced while the model writes, and think turns a reasoning model's thinking on or off. RoachInference reads text only: since 1.1.3, a request with images gets a 400 that says so.

/api/create makes a new model over an installed one or a local .gguf file, from fields or a Modelfile (FROM, SYSTEM, PARAMETER, MESSAGE). The weights are hard-linked, not copied. /api/tags lists every model the server can load, including models read in place from an Ollama models folder.

OpenAI routes

/v1/chat/completions takes tools, tool_choice, response_format (json_object or json_schema), reasoning_effort and stream_options.include_usage, and streams server-sent events with "stream": true. /v1/responses keeps nothing between calls, so send the history you want each time. Since 1.1.3, an image, audio, video or file part gets a 400, and an embeddings input that is neither text nor a list of texts is refused.

Anthropic routes

/v1/messages takes text, tool_use and tool_result blocks (an image or document block, in a message or a tool result, gets a 400 since 1.1.3), tools, tool_choice and thinking, and streams Anthropic's event stream with "stream": true. /v1/messages/count_tokens counts with the model's own tokenizer and template.

Every route

17 routes, from the route data the route browser uses, checked against Router.swift at tag v1.1.4.

Related

• Use RoachNet with Ollama, OpenAI and Anthropic clients has examples for each format.
• RoachClaw covers which models load and how memory is planned.