Confab is a small family of AI tools built around the same idea: your models, your keys, your machine. Pick a native macOS app or Jarvis ‐ a self-hosted system on the same prinicpal as Confab but with even more tools. Nothing routes through us, nothing leaves your infrastructure.
Every AI, one interface. Confab connects OpenAI, Anthropic, OpenRouter, Google, and local Ollama models into a single, beautifully native macOS experience ‐ with smart memory, file access, and Git sync built in.
OpenAI, Anthropic, OpenRouter (300+ models), Gemini, and local Ollama ‐ switch with one shortcut, no context lost.
Cut token costs 50–80% with intelligent context selection, message starring, and progressive summarization.
Grant read (and optionally write) access to specific folders. Ask questions about your own code and documents.
Let the AI run shell commands via /bin/zsh, opt-in with a per-command or session-wide approval gate.
Conversations export as Markdown and push to your Git repo automatically ‐ free backup, zero infrastructure.
One-click backup of settings to iCloud Drive. API keys are excluded and must be re-entered after restore.
A self-hosted personal AI application that runs entirely on your own hardware. Connect OpenRouter, Claude, or OpenAI models to your calendar, email, files, and the web or create you own short running, long running or periodical running agents controlled with cron ‐ with every permission under your explicit control.
Copy docker-compose.example.yml, .env.example, SOUL.md.example, and USER.md.example into a server directory and drop the .example suffix.
mkdir -p data ‐ this is where Jarvis persists its database and files.
Edit .env and fill in the required (*) values: a master password for the credential store, the database connection strings, and an MCP key generated with openssl rand -hex 32.
One command pulls the images and starts Jarvis and PostgreSQL.
WebSocket streaming with full tool-call visibility. Watch Jarvis reason and act, switch models mid-conversation.
Read/write events via CalDAV, manage inbox via IMAP/SMTP ‐ zero data leaving your server.
Read, create, and edit files within explicit directory boundaries you declare.
PostgreSQL + pgvector knowledge base. Capture thoughts, search by meaning, build long-term memory.
Goal-oriented agents with their own tools, models, and cron schedules. Can spawn sub-agents.
Prompt injection screening, canary tokens, append-only audit log, explicit confirmation on destructive actions.
Stream responses from Claude or any OpenAI-compatible model. See tool calls as they happen, confirm or deny actions inline, and switch models per conversation.
Agent Skills are markdown instruction files that teach the AI how to behave. Active skills are automatically injected into the system prompt for every conversation ‐ the AI reads them and applies the instructions when relevant.
Skills follow the SKILL.md open standard, making them compatible with a growing
ecosystem of skill marketplaces and community collections.
More skills coming soon.
Questions, bug reports, or feature ideas ‐ send a message and we’ll get back to you.