Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS
Grok Bot box — free local Persistent Mind
Audience: PortOS installs running on a Grok Bot box (CPU-only, ~16 GB RAM, no GPU) and similar modest hosts.
Recommended topology
| Role | Where it runs | Default |
|---|---|---|
| Persistent Mind | Local Ollama API provider | Qwen2.5 7B Instruct (Q4) — qwen2.5:7b-instruct (alias: qwen2.5:7b) |
| Coding tasks | Cloud CLIs / harnesses | Cursor Agent / OpenCode Zen (and other subscription CLIs you enable) |
Keep the mind free and local. Do not point coding agents at a CPU-bound 27B/vLLM stack on this host — those presets need a GPU workstation (see Recommended coding-agent setup, SGLang, Fleet LLM host).
Why this default
- Persistent Mind is an always-on conversational agent with optional tool grants. A tool-capable ~7B instruct model fits ~16 GB RAM at Q4 and stays responsive enough for wake turns.
- Cursor / OpenCode Zen already own the coding harness (worktrees, tools, PRs). Using them for coding avoids pretending a CPU box is a 27B coding workstation.
- The built-in
ollamaprovider is a Direct API lane — the reliability path Persistent Mind prefers over TUI scraping.
Setup UI
PortOS surfaces a Free local Persistent Mind card when the host fits this path (CPU-only / no usable NVIDIA GPU, or low RAM / modest Apple Silicon without a curated 48 GB+ coding preset):
- AI Providers (
/ai) - Models → Runtimes (
/models/llms-runtimes) - Chief of Staff → Persistent Mind → Settings (AI profile drawer)
The card:
- Detects whether Ollama is installed and answering.
- Offers Install Ollama / Start Ollama / Pull Qwen2.5 7B Instruct via the existing local-LLM APIs.
- Enables the built-in
ollamaprovider and pins the recommended model. - Optionally sets the Persistent Mind profile to that provider/model.
Copy on the card states that this is the default for Grok-box / CPU-only / no-GPU hosts, and warns against 27B / vLLM presets.
API
| Method | Path | Purpose |
|---|---|---|
GET |
/api/local-llm/persistent-mind-setup |
Checklist + recommendation (or applicable: false) |
POST |
/api/local-llm/persistent-mind-setup/apply |
Enable ollama + pin model; optional { setMindProfile: true } |
Apply never downloads weights. Install / start / pull stay on /api/local-llm/install-backend, /ollama-service, and /install.
Guardrails
- No GPU / low RAM: the recommendation includes explicit warnings; expect slower CPU inference.
- Heavy presets suppressed: hosts on this path should not be steered to Qwen3.8-27B, vLLM, or SGLang as the default. Curated GPU coding hosts keep Hardware LLM recommendation instead and do not get this card.
- Hardware gates on catalog entries (e.g. 27B
minMemoryGb: 32, vLLMrequiresNvidiaGpu) still hide unavailable options in the library.
Related
- Claude Ollama — local model behind the Claude Code harness (optional coding path once Ollama is up)
- Persistent Mind / CoS enhancement
- Product surfaces