nc://ollama-cloud

A Nodecode cell for Ollama Cloud, Ollama's hosted runtime. With it, /connect offers Ollama Cloud, /models lists its models (asked of GET /api/tags), and a turn on ollama-cloud/<model> goes to https://ollama.com/api/chat over Ollama's own wire, the way oh-my-pi sends it.

Why a lane of its own

models.dev lists ollama-cloud on the OpenAI-compatible chat wire (https://ollama.com/v1), which the core's chat lane already speaks. omp does not use it: its catalog maps that very models.dev row to its ollama-chat api (provider-models/openai-compat.ts, simpleModelsDevDescriptor("ollama-cloud", "ollama-cloud", "ollama-chat", "https://ollama.com")) and serves all 52 bundled models over POST /api/chat. omp's source states no single reason; what the native wire carries that /v1 does not is:

• the think value, with Ollama's own rungs low, medium, high and max; the DeepSeek V4 and GLM 5.2 ladders end at max, which no reasoning_effort spelling reaches, and think: false turns thinking off
• the prompt's cache split (prompt_eval_cached_count), so cache reads are counted apart from the rest of the prompt
• done_reason: "load", Ollama's answer to a request with no user turn, which omp turns into an error instead of an empty answer
• a tool result named by its tool (tool_name), images as base64 arrays

The cell ports omp's choice: it registers a lane, ollama-cloud, whose stream (wire.lisp) speaks /api/chat:

• the request is {"model", "messages", "tools"?, "think"?, "tool_choice"?, "options"?: {temperature, top_p}, "stream": true}; the system prompt leads as a system message, a history system message (an eviction stub) stays system, a tool result is {"role": "tool", "tool_name"}, an assistant's calls go back with their arguments as objects, and its thinking is never sent back (Ollama Cloud answers 400 to history that carries thinking)
• a request with no user turn gets one: the last system turn past the prompt's own becomes user
• an image rides as base64 in images for a model that takes images; for one that does not, [image omitted: model does not support vision] stands in its place
• the effort maps to think as omp maps it: off is false for a thinking model, minimal and low are low, xhigh is high, max is max; no effort leaves the key off
• tool_choice is none or required (auto is left off); tool schemas pass omp's Ollama sanitizer: a boolean subschema widens to a union of every primitive type, a boolean additionalProperties goes, a type list folds to one type or an anyOf under allOf
• num_predict is never sent: omp marks every Ollama Cloud model omitMaxOutputTokens, so the output ceiling is the endpoint's
• the answer is NDJSON, one chunk per line: message.thinking is reasoning, message.content is text, message.tool_calls arrive whole (each named ollama:<block>:<name>, omp's id), and the done chunk carries the finish and the usage (prompt_eval_count less its cached part as input, the cached part as cache reads, eval_count as output)
• a round with calls is a tool round whatever done_reason says; length with nothing in it is the context overflow the core evicts on; load is an error; an {"error"} line on a 200 is said as the provider's error; a stream that ends without done is a truncated one

Ported from oh-my-pi's ollama-cloud provider. See NOTICE.

Key

Make a key at <https://ollama.com/settings/keys>. Save it with /connect, or set OLLAMA_CLOUD_API_KEY. A saved key wins over the variable. /connect checks a typed key by asking GET /api/tags with it: a 401 is a refused key, any listing an accepted one. No other provider's key (OPENAI_API_KEY among them) is ever sent to Ollama.

Install

Copy this folder into ~/.nodecode/cells/nodecode-ollama-cloud/ and restart Nodecode, or run (restart-cells).

Configure

Optional. The section is on unless it says "enabled": false:

{
  "ollama-cloud": {
    // where Ollama's native API is served; the lane appends /api/chat
    "base_url": "https://ollama.com"
  }
}

To keep the OpenAI-compatible wire instead, pin it the core's way, which the cell leaves alone (the key still comes from this cell):

{
  "providers": {
    "ollama-cloud": {"sdk": "openai-completions", "base_url": "https://ollama.com/v1"}
  }
}

Test

From this repository's root:

tools/test-cell.sh nodecode-ollama-cloud

Not ported

• omp's stream markup healing (utils/stream-markup-healing.ts and the dialect/ scanners behind it): Kimi's <|tool_calls_section_begin|> tokens and DeepSeek's DSML envelope leaked into content are not turned back into tool calls, and leaked <think> idioms are not lifted out of the text. The core's own route still takes a leading <thinking> block as reasoning.
• omp's message normalization (transform-messages.ts, 1,300 lines: cross-model thinking rewrites, orphan repair); the core's history fold already pairs every call with a result.
• omp's own retry of a 5xx (2 s, 5 s, 10 s, except llama.cpp's tool-call parse failure) and its replay of an empty answer: the core's retry policy and its PROVIDER-EMPTY-RESPONSE retry stand in for both.
• the per-model effortMap (no bundled row carries one) and the first-event watchdog (the core's idle deadline covers the wait for headers and every line after).
• Discovery asks /api/show only for a model the bundled rows do not carry, for its window; omp asks it for every model, for its window and its thinking and vision capabilities. A listed model the rows do not carry is taken to think (an effort reaches it as think) and to take no images.

MIT licensed.