ollama

name	ollama
description	Calls Ollama Cloud for GLM 5.2 at high reasoning, then executes the returned plan. Use when the user types /ollama, says Ollama Cloud, cloud GLM, or asks to plan/execute via Ollama Cloud GLM 5.2. Not for local GPU Ollama (use ollama-local-setup). Never localhost:11434 for this chair. Never print OLLAMA_API_KEY.
disable-model-invocation	true

/ollama — Ollama Cloud GLM 5.2

One capability: Ollama Cloud API as planner, then execute in this session.

This is not local GPU Ollama. Local install/run belongs to ollama-local-setup.

Official Cloud docs: https://docs.ollama.com/cloud

When to Use

• User runs /ollama <task> or says Ollama Cloud / cloud GLM 5.2.
• User wants a Cloud GLM plan, then this agent to execute it.

Do not use for:

• Local models on http://localhost:11434 — ollama-local-setup (and /local if that rule is installed).
• Robotics, hardware bring-up, or generic “run any LLM” tasks.
• Substituting Cloudflare GLM MCP when the user asked for /ollama.

Prerequisites

• Env OLLAMA_API_KEY already set (create at https://ollama.com/settings/keys). Never print the key.
• Network to https://ollama.com.
• Optional helper on disk: $env:LOCALAPPDATA\hermes\scripts\ollama_cloud_api.py — use if present; do not require a machine-specific path.

API contract

Setting	Value	Source
Native Cloud host	https://ollama.com	docs.ollama.com/cloud
Native chat	POST https://ollama.com/api/chat	same
OpenAI-compatible	POST https://ollama.com/v1/chat/completions	Ollama as remote host + OpenAI compat
Auth	Authorization: Bearer + env OLLAMA_API_KEY	docs.ollama.com/api/authentication
Default model	glm-5.2 (no :cloud suffix on the Cloud host)	this chair
Reasoning	reasoning_effort: "high"	OpenAI-compat field; omit max_tokens unless the user caps

Procedure

1. Parse the task after /ollama. If missing, ask once.
2. Gather minimal context — file paths, not huge dumps.
3. Call Cloud (prefer the helper if the file exists; else curl). Do not print the key. Do not use localhost:11434. Do not use ollama run glm-5.2:cloud as a substitute for this chair.
4. Execute the returned steps (adapt for safety, project rules, PowerShell).
5. Verify with the project’s checks or the verification GLM specified.
6. Report what GLM decided, what ran, evidence, blockers.

Native Cloud call (documented)

curl.exe -sS https://ollama.com/api/chat `
  -H "Authorization: Bearer $env:OLLAMA_API_KEY" `
  -H "Content-Type: application/json" `
  -d '{"model":"glm-5.2","stream":false,"messages":[{"role":"user","content":"<TASK>"}]}'

OpenAI-compatible call (remote host)

curl.exe -sS https://ollama.com/v1/chat/completions `
  -H "Authorization: Bearer $env:OLLAMA_API_KEY" `
  -H "Content-Type: application/json" `
  -d '{"model":"glm-5.2","reasoning_effort":"high","messages":[{"role":"user","content":"<TASK>"}]}'

Optional helper

$helper = Join-Path $env:LOCALAPPDATA "hermes\scripts\ollama_cloud_api.py"
if (Test-Path $helper) {
  python $helper agent "<TASK>" --reasoning high
}

Modes

User says	Behavior
/ollama … (default)	Plan via Cloud then execute
/ollama plan … or “plan only”	Call Cloud, present plan, stop
/ollama <other-model> …	Use that Cloud model id if listed by GET https://ollama.com/api/tags; else glm-5.2

Anti-patterns

• http://localhost:11434 for /ollama
• ollama run glm-5.2:cloud instead of the API-key Cloud host
• Cloudflare ask_glm_5_2 when the user asked /ollama
• Echoing or embedding OLLAMA_API_KEY
• Passing max_tokens by default
• Stopping after the plan when the user wanted the job done
• Robotics / TODO boilerplate — out of scope

Related

• ollama-local-setup — local GPU Ollama on localhost:11434 (install, verify script).
• Official Cloud: https://docs.ollama.com/cloud
• Model card: https://ollama.com/library/glm-5.2