ConductLoop
Local-first second brain and AI-agent cockpit. Python + SQLite backend, native macOS app (Tauri), bring-your-own-key models.
ConductLoop indexes your own material — an Obsidian vault, memory/notes folders, code repositories, a skills directory, project playbooks and a list of URLs you want to watch — into a single local SQLite database with full-text search (FTS5), fuzzy matching and optional local semantic embeddings. On top of that index it runs a Brain (RAG chat with citations) and a set of AI agents/teams (Boardroom, tasks, runs, scheduler) using the model provider you configure.
The Python package is still named ops_dashboard for backwards compatibility. The product name is ConductLoop.
Principles
Local-first
- The server is a stdlib
http.server that binds to 127.0.0.1:7777 by default.
- All indexed content, memory, tasks and settings live in a local SQLite file (
ops.db). Location, in order of precedence:
$OPS_DATA_DIR/ops.db<repo>/data/ops.db (dev checkout, if data/ exists)- macOS
~/Library/Application Support/com.conductloop.desktop/data/ops.db · Linux ~/.local/share/conductloop/ops.db · Windows %LOCALAPPDATA%\ConductLoop\data\ops.db
- At-rest secrets (DB encryption key, connector credentials) go to the macOS Keychain, or to
0600 files in the data dir elsewhere (ops_dashboard/secrets_store.py). Connector credentials are Fernet-encrypted in the DB.
- Network calls only happen for features you turn on: your model provider, connectors you configure (GitHub, Slack, Telegram, Discord, Google Search Console/Analytics, Sentry, …), URL health checks, webhooks, and the Tauri auto-updater in the desktop app.
- Error reporting to Sentry is opt-in (only if you set
SENTRY_DSN).
- If you install
sentence-transformers, the embedding model (all-MiniLM-L6-v2) is downloaded once from Hugging Face and then runs locally.
BYOK (bring your own key)
The Brain talks directly to the provider you configure; nothing is proxied through ConductLoop servers.
Provider (BRAIN_PROVIDER) | Needs |
|---|
anthropic | ANTHROPIC_API_KEY |
openai | OPENAI_API_KEY |
minimax | MINIMAX_API_KEY (OpenAI-compatible; MINIMAX_BASE_URL optional) |
gateway | GATEWAY_API_KEY for an OpenAI-compatible gateway (GATEWAY_BASE_URL, default Vercel AI Gateway) |
ollama | a local Ollama on localhost:11434 (OLLAMA_MODEL, OLLAMA_BASE_URL optional) |
mock | nothing — offline/testing |
BRAIN_PROVIDER=auto (default) picks the first available provider; BRAIN_MODEL overrides the model. Keys can also be set from the UI (/credentials), which stores them in the macOS keychain or, where there is none, in an encrypted file protected by a passphrase — never in plain text. Other settings live in your local config.json ($OPS_CONFIG_PATH, or in the data directory); it is gitignored — never commit it.
Install
macOS app
Download the latest .dmg from conductloop-releases, drag ConductLoop to /Applications and open it. Requires macOS 13 (Ventura) or later on Apple Silicon; the built-in browser needs macOS 14 (Sonoma) or later (on macOS 13 everything else works). The app is a Tauri v2 shell that starts the bundled Python server as a sidecar and shows it in a native WebView.
From source (macOS / Linux)
Requires Python 3.13+ (CI runs 3.14).
git clone https://github.com/EVMCOD/conductloop.git
cd conductloop
python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
make serve # → http://127.0.0.1:7777
Then, optionally, point it at your content and index it:
cp .env.example .env # the make targets load .env automatically
make index # index all configured sources
Dependency sets:
requirements-core.txt — lean runtime (server, CLI, auth, search). No ML.requirements.txt — core + Anthropic/OpenAI SDKs + local semantic search: fastembed (ONNX, no PyTorch; what the app uses) and sentence-transformers (pulls PyTorch, ~2 GB).requirements-optional.txt — Whisper voice notes, Google Analytics, PyJWT, Sentry, MCP server.requirements-dev.txt — core + pytest, ruff, mypy, pip-audit, pre-commit.
pip install -e . also installs the ops CLI entry point.
Configuration
A fresh install is empty: no projects, teams or URLs. Point it at your content with environment variables (see .env.example), e.g. OPS_VAULT_ROOT, OPS_WORKSPACE, OPS_SKILLS_DIR, OPS_REPOS, OPS_PROJECTS_FILE, OPS_LIVE_URLS_FILE. Set CONDUCTLOOP_DEMO=1 (with a throwaway CONDUCTLOOP_DB_PATH and OPS_DATA_DIR) to load fictional demo data instead; see docs/FIRST-5-MINUTES.md.
Optional notification webhooks (Slack / Discord / Telegram) are described in config.example.json. Remote access from a phone is documented in docs/REMOTE.md (Tailscale; nothing is exposed publicly by default).
Usage
make help # list targets
make index # reindex everything
make serve # start the web UI on :7777
make brain # ask the Brain from the terminal
make briefing # morning briefing
make snapshot # take a daily snapshot
make diff # diff today vs yesterday
Selected ops / python -m ops_dashboard commands: index, find, brain, briefing, capture, todos, projects, playbook, snapshot, backup, export / import, sync-export / sync-import (passphrase-encrypted bundle), obsidian-sync, calendar-import, mcp (expose the Brain over MCP), setup (readiness checklist). Run ops --help for the full list.
Development
python3 -m venv .venv && . .venv/bin/activate
pip install -r requirements-dev.txt
pre-commit install # optional: ruff + basic hygiene hooks
ruff check ops_dashboard tests # lint (blocking in CI)
mypy ops_dashboard # type check (advisory in CI; a strict subset is blocking)
pytest -m "not live" # hermetic test suite (what CI runs)
Tests marked live need the network or real credentials; they never run by default or in CI.
Layout
ops_dashboard/ Python package: server.py (HTTP API + UI), cli.py (typer CLI),
db*.py (SQLite schema/queries), brain.py (RAG + providers),
indexers/ (vault, memory, repos, skills, urls, …), web/ (vanilla HTML/JS UI)
desktop/ Tauri v2 macOS app (Rust shell + PyInstaller sidecar build)
desktop-demo/ Tauri demo build
scripts/, bin/ launchd jobs, release and maintenance scripts
tests/ pytest suite
Desktop app
See desktop/README.md and desktop/docs/DISTRIBUTION.md. Building and signing a release needs your own Apple Developer ID and your own Tauri updater key; neither is in this repository.
Linux (experimental) and Windows are not shipped. docs/PORTABILITY.md is the audit of everything that is macOS-only today, with the plan to port it.
Contributing & security
License
MIT. The ConductLoop name and logo are not covered by the license.
