Ollama Delegate

Runs tasks through the Ollama HTTP API (/api/generate for one-shot, /api/chat for multi-turn) rather than shelling out interactively. Works against http://localhost:11434 (the default local Ollama daemon) or any other reachable Ollama-compatible API base URL the user gives you — including an external/remote Ollama instance.

See also: README.md for the API mechanics behind this design, USE_CASES.md for trigger phrases and worked examples, and the top-level skills index for how this fits alongside codex-delegate.

When to use this

Use it for: a task suited to a smaller/local model that doesn't need Claude-level reasoning — summarization, classification, extraction, translation, bulk drafting, format conversion, or a first pass on many similar items. Also use it when the user explicitly asks for Ollama or a local/self-hosted model.

Don't use it for: tasks that need deep reasoning, multi-file code changes, or tool use — Ollama models here are called through a plain HTTP API with no file access, no shell, and no tools of their own. If the task needs to read/write real files or run commands, do that yourself or use codex-delegate, then optionally hand only the text content to Ollama.

This skill parallelizes. Ollama's daemon can serve multiple requests concurrently (subject to its own queuing/OLLAMA_NUM_PARALLEL config), so unlike codex-delegate's strict one-at-a-time rule, you may fire off several independent run calls — e.g. one per file, one per item in a batch — at once. Still keep each multi-turn start/send thread sequential (a single state file is one conversation; don't send to it from two calls at once).

Prerequisites: resolving host and model

Do this once per task, in order, before any run/start call: