Going Fully Local

journal works entirely on your machine by default — your notes are never uploaded for indexing or search. But if you also want synthesis (AI summaries) to run locally, with zero data leaving your computer at all, you can set that up too.

This guide walks through a complete zero-egress setup using Ollama for both embeddings and synthesis.

What "fully local" means

By default, journal synth sends your note excerpts to Anthropic's cloud API. The local_only setup replaces that with a local Ollama model, so:

• Indexing: local (already the default)
• Search: local (already the default)
• Synthesis: local (this is what we're adding)
• MCP tools: optionally local, if your chat client runs locally too

Nothing leaves your machine.

Hardware requirements

Ollama loads models on demand and unloads them after about 5 minutes of inactivity, so these are transient peaks, not standing costs.

Step 1: Install and start Ollama

macOS:

brew install ollama
brew services start ollama

Linux:

curl -fsSL https://ollama.com/install.sh | sh
# Installed as a systemd service, starts automatically

Verify it's running: ollama --version

Step 2: Pull both models

ollama pull qwen3-embedding:4b   # for indexing and search (~2.5 GB)
ollama pull gemma4:12b           # for synthesis (~8 GB, or pick a smaller model)

Step 3: Update your config

In your journal repo's .journal/config.yaml:

synth_provider: ollama
synth_ollama_model: gemma4:12b

local_only: true          # block all cloud AI paths
local_only_mcp: allow     # if you want to use journal with a local MCP chat client

The local_only: true flag is a hard kill-switch — it refuses cloud synthesis and requires Ollama to be on loopback. It doesn't affect journal sync, which backs up to your own git remote (not cloud AI).

local_only_mcp: allow is only needed if you plan to use journal mcp with a local chat client like Jan or LM Studio. Leave it as block (the default) until you've set up the client.

Step 4: Verify

journal doctor

The egress line should say something like:

local_only: no cloud-AI egress (synth local: gemma4:12b); mcp blocked by policy

Try it:

journal synth daily              # dry run: shows the prompt, no network call
journal synth daily --write      # runs locally, writes reflections/daily-YYYY-MM-DD.md
journal search "what did I work on"   # semantic search, fully local

Optional: Enable the reranker

The reranker is a local quality boost for search results. Without it, results are in vector-distance order — already high quality for most queries. With it, the top vector-KNN candidates are re-scored by a small generate model before returning, giving a precision lift especially on longer or ambiguous queries.

Pull a small generate model and add one line to your config:

ollama pull qwen3:4b   # ~2.5 GB — the recommended reranker model

In .journal/config.yaml:

reranker: qwen3:4b

That's it. journal doctor confirms the reranker is configured. It runs on the same loopback Ollama — no network egress. The extra RAM cost is transient (~2.5 GB while reranking; Ollama unloads the model after ~5 min of inactivity).

The reranker is off by default. qwen3-embedding:4b is strong enough for most queries on its own; the reranker matters most when you're getting subtly irrelevant hits and want to tighten precision.

Adding a local MCP chat client (optional)

If you want to chat with your notes using a local AI model (no Claude), you can pair journal's MCP server with Jan or LM Studio. Both can connect to Ollama and call journal tools.

See Local MCP Clients for setup instructions for each client.

Quality expectations

For summarization tasks — daily digests, weekly rollups, decision rollups — local models like gemma4:12b perform very close to cloud Claude. The gap shows most on long-form stylistic writing where voice profile matching matters a lot.

A practical approach: run daily synthesis and search answers locally, and keep weekly digests on synth_provider: anthropic if you care about the writing quality. You can switch providers with a single line in config.