Fully Local Ollama Agent (No Cloud)

This playbook builds an agent that runs entirely on a local Ollama provider through the Pi SDK runtime, with journal memory backed by local embeddings and a sandbox that blocks all non-localhost network traffic. No cloud API keys, no outbound calls.

Who this is for

A privacy-focused user with no cloud API budget who wants the whole stack — chat model, embeddings, and recall — to stay on their own machine.

Goal

An agent that runs entirely on a local Ollama provider via the Pi SDK runtime, with journal memory using local embeddings and no outbound network.

Features used

• runtime.local-providers — register Ollama and reference it as ollama:<model>
• runtime.multi-backend — the Pi SDK bridge that drives the local provider
• memory.journal — hybrid (BM25 + vector) recall with static, non-decaying salience, no LLM required
• memory.embeddings-config — local Ollama embeddings for the journal tier
• sandbox.network-policy — restrict runtime commands to localhost

Configuration

Every key below is verified against the config blueprint. The providers map registers Ollama; runtime.model references it as ollama:llama3.1:8b. Journal memory needs an embeddings block, and sandbox.network.mode: "localhost" keeps all runtime command traffic on-box.

{
  "runtime": {
    "model": "ollama:llama3.1:8b"
  },
  "providers": {
    "ollama": {
      "type": "ollama",
      "baseUrl": "http://localhost:11434",
      "enabled": true,
      "models": [
        {
          "name": "llama3.1:8b",
          "capabilities": { "context_window": 32768 }
        }
      ]
    }
  },
  "memory": {
    "mode": "journal",
    "path": "./.mono-agent/memory",
    "embeddings": {
      "provider": "ollama",
      "model": "nomic-embed-text:v1.5",
      "endpoint": "http://localhost:11434",
      "dim": 768
    }
  },
  "sandbox": {
    "mode": "native",
    "network": { "mode": "localhost" }
  }
}

Caution

Use the exact nomic-embed-text:v1.5 tag — the bare nomic-embed-text name resolves to a different revision and will not match the stored 768-dim vectors.

Caution

Model size drives first-turn latency. A light 8B model (the local-ollama-private recipe default) keeps the first turn in seconds-to-minutes on typical hardware. A large model — e.g. a 30B/Q4 loaded at its full 128k–256k context window — makes Ollama allocate a huge KV cache; measured first turns ran 274s to 570s+ before responding. Pi drives Ollama over the OpenAI-compatible endpoint, which has no num_ctx knob, so the reliable lever is model choice, not a context cap. Size your machine before pointing --model at anything heavier.

Note

The journal tier is config-only and needs no memory.llm; recall is embeddings + FTS with no model call. Promote to mode: "bujo" later if you want LLM capture and the scheduled bujo-tier consolidation path, which requires a memory.llm block for effective tier selection. See Capture and recall.

Steps

1. Pull both models locally: ollama pull llama3.1:8b && ollama pull nomic-embed-text:v1.5
2. Scaffold the agent: mono-agent init --model ollama:llama3.1:8b --memory journal
3. Add the providers entry for Ollama and the memory.embeddings block, then set sandbox.network.mode to localhost.
4. Run mono-agent validate to confirm Ollama is reachable and both models are pulled.
5. Run mono-agent start — keep the webhook channel as the zero-credential smoke channel.
6. curl the webhook endpoint and confirm a response generated by the local model.

Smoke test

Tip

curl -X POST the webhook invoke path with a prompt; verify a response generated by the local Ollama model and that the run artifact shows no outbound non-localhost network.

Related

• Local providers
• Runtime backends
• Embeddings
• Capture and recall
• Sandbox
• Webhook channel
• Artifacts and traces — confirm no outbound network in the run artifact
• Skills and the mono-agent-composer — scaffold this config end to end