Minion
Minion
Minion is a capture-first macOS second brain: it watches what you work on (window focus, accessibility text, selective screenshots when text is thin), ingests files you drop in, and stores everything in a local vault at ~/Library/Application Support/Minion/data. The app opens to Activity—one Slack-like stream of capture, sync, parsing, council proposals, and 42 conversation—not a search box you have to remember to use. 42 is the proactive guide that asks questions and offers guidance; you answer in short bursts. Claude Desktop (and other MCP clients) can read the same vault over stdio when you want an assistant in the loop.
What you get
| Surface | Role |
|---|---|
| Activity | Home: chronological river of context—ambient capture, ingest/sync jobs, health issues, council suggestions, and 42 threads. |
| 42 | Proactive guide embedded in Activity. Surfaces questions and nudges so you do not have to invent good prompts; reply or dismiss. |
| Sources | Drop zone, inbox watcher, ingest toggles, and search over your indexed corpus. |
| Mirror / graph | Life-graph scaffold (people, projects, obligations, …) with evidence attached; identity mirror and wiki routes where enabled. |
| MCP | Memory tools for Claude (ask_minion, working context, voice helpers, wiki/work proposals)—local index only, no hosted memory service. |
Under the hood: one memory.db (plus vectors) in the vault; override the data directory with MINION_DATA_DIR. macOS capture respects deny lists and keeps screenshots sparse (OCR when accessibility text is thin). Settings cover sidecar paths, Claude one-click MCP setup, and per-format ingest.
Ships today: Tauri desktop, Python sidecar, ambient capture, Activity + 42, vault + MCP, life-graph seed, council-style suggestions. Roadmap (mirror sharing, belief plasticity, tiered compression at scale, ambient audio, and more) is spelled out in docs/ROADMAP.md—not promised here.
Documentation
| Doc | Contents |
|---|---|
docs/SECOND_BRAIN.md |
Four-loop model, routes, Activity feed, MCP posture |
docs/LIFE_GRAPH.md |
Graph scaffold, domains, retrieval order |
docs/SCREEN_MEMORY.md |
Screen-memory capture, fuse, retrieval, and guidance loop |
docs/SCREEN_ADAPTERS.md |
Playwright, Marlin, and OmniParser adapter command setup |
docs/SCREEN_MEMORY_AUDIT.md |
Requirement-to-evidence checklist for the screen-memory MVP |
docs/ROADMAP.md |
Product direction and phased themes |
chatgpt_mcp_memory/README.md |
Tool tables, parsers, env flags, CLI |
desktop/README.md |
Tauri architecture, tauri dev / release bundles |
AGENTS.md · PROCESS.md |
Agent playbook and delivery loop (contributors) |
Repo layout
| Path | What it is |
|---|---|
desktop/ |
macOS app: SvelteKit UI + Tauri 2 Rust shell (src-tauri/). |
chatgpt_mcp_memory/ |
Python sidecar (FastAPI, ingest, SQLite, MCP). |
On tauri build, desktop/src-tauri/scripts/sync_sidecar.sh copies chatgpt_mcp_memory into desktop/src-tauri/resources/sidecar/ (gitignored; canonical source stays chatgpt_mcp_memory/).
Context makes the difference
When an AI has access to Minion's index, the output changes character entirely. Here Claude is asked to rewrite website copy "in my voice — ask minion". It calls ask_minion across dozens of past conversations, reads the pattern ("short sentences, direct subject-verb, 'we' naturally, thinks out loud without hedging"), and rewrites the copy in that register — quoting actual lines back as evidence.
Without Minion: generic, polished, forgettable. With Minion: the AI knows how you actually write and produces something that sounds like you wrote it yourself.
Screenshots
| Main window — drop zone and ingest | Activity — live parse / embed log |
|---|---|
| Claude Desktop — Minion MCP in a real thread | macOS — Minion in Launchpad |
|---|---|
Preferences — Status (sources, chunks, sidecar, paths), Claude (MCP) one-click config, and Ingest & file types:
| Status | Claude (MCP) | Ingest & types |
|---|---|---|
Install (macOS app)
1. Pick the right download (Intel vs Apple Silicon)
On GitHub Releases there are two macOS downloads. Read the file name—it says which kind of Mac it is for:
| Download file name contains… | Use for | How to tell on the Mac |
|---|---|---|
…-macOS-Apple-Silicon.zip |
Apple Silicon — M1, M2, M3, M4, … | Apple menu → About This Mac → Chip: “Apple M2” (or any Apple M…). |
…-macOS-Intel.zip |
Intel — Core i5 / i7 / i9, Xeon, … | About This Mac → Processor: line includes Intel. |
If you pick the wrong one, macOS may refuse to open the app or show an architecture error. Delete Minion.app, download the other zip, unzip, and drag Minion.app to Applications again.
2. Install and run
- Download the matching zip from GitHub Releases (or build from source — below).
- Unzip, then move Minion.app to Applications when prompted (avoid running forever from the disk image or a translocated copy).
- First launch can take a few minutes while Minion prepares its embedded Python environment and starts the sidecar. Later launches are quick.
- Open Activity to see capture and ingest; engage 42 when it surfaces a thread. Optional: open Settings → Claude (MCP) and click Add to Claude, then fully quit and reopen Claude Desktop.
- Optional: import a ChatGPT export (zip or folder) via the drop zone so the index has history on day one.
Ollama (image captions): If you do not already have Ollama installed, Minion can download the official macOS Ollama app into your Minion data folder on first launch and start it locally. To skip that behavior, set MINION_SKIP_MANAGED_OLLAMA=1 before opening the app.
Build from source (developers)
git clone https://github.com/reif-is-a-foofie/Minion.git
cd Minion # or your clone path, e.g. ~/Classified/minion
git submodule update --init --recursive
cd chatgpt_mcp_memory
python3.11 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cd ../desktop
npm install
npm run tauri dev
The Rust shell prefers ../chatgpt_mcp_memory/.venv/bin/python. Release builds produce desktop/src-tauri/target/release/bundle/macos/Minion.app; see desktop/README.md for bundling notes.
CLI (minion command)
For a terminal-first setup (export path, minion doctor, minion setup, inbox CRUD without the GUI), use the launcher in bin/minion and chatgpt_mcp_memory/README.md. The desktop app and the CLI share the same store and MCP server code.
Screen-memory MVP commands:
./bin/minion remember-screen
./bin/minion search "where did I see the export button?"
./bin/minion summarize-last 30m
./bin/minion what-was-i-doing
./bin/minion guidance
./bin/minion screen-memory-status --probe
./bin/minion verify-screen-memory
When run from this checkout, screen-memory commands use the repo-local chatgpt_mcp_memory/.venv automatically. Use --workspace only when you want to run against another installed workspace.
Privacy
Indexing, capture, and search run on your machine. MCP speaks stdio to Claude Desktop by default—no cloud “memory service” for your chunks. Optional components (for example Ollama for voice or captions, or HF Hub for some embedding downloads) only touch the network if you configure them. Keep exports and large corpora outside the git tree if you prefer.
Credits
Built and dogfooded by Reif — questions: [email protected].
Cursor skills catalog: Awesome Cursor Skills by Spencer Pauly — vendored in-repo at third_party/awesome-cursor-skills. Third-party list; not affiliated with Minion.
