Babo

Self-hosted agent OS — persistent memory, projects & teams, WhatsApp · Telegram · Google.
MIT · local-first · no telemetry · babo.agency

Download desktop · Manifesto · Docs · Discord

A persistent AI agent that works with you — across chat, projects, channels, and tools.

Babo is a full-stack agent platform: a desktop app, web UI, and self-hostable backend. Each agent has long-term memory, an autonomous agentic loop, project orchestration with sub-agents, and one-click integrations for the channels and tools you already use.

Connect any OpenAI-compatible model (OpenRouter, Ollama, local vLLM, LAN server, Babo Cloud hybrid routing, etc.) and run Babo on your machine or your own server.

Audiences: Builders (OSS, self-host) vs Everyone (personal AI, no terminal). See docs/audiences.md and PRODUCT.md.

Not just a chat UI — Babo is an extensible agent platform: native skills, agent tools, and channels you can add in code (MCP, ClawHub, crystallize). Memory, board, and teams survive sessions. See the manifesto and extension guide.

---

What Babo can do

Chat with a real agent

• Streaming conversations with live tool use, thoughts, and file proposals
• Agentic loop — the agent orients, plans, calls tools, evaluates progress, and keeps going until the job is done
• Structured plans — multi-step runbooks with sub-plans, acceptance criteria, and verification
• Slash commands — e.g. trigger a sleep cycle to consolidate memory on demand
• Voice input — speak to your agent (local or remote transcription)
• Embedded browser — watch the agent browse and work in its own workspace

Projects, board, timeline & teams

The Projects workspace is where Babo runs serious work:

Babo can spawn sub-agent teams to work in parallel on plan steps. You see active teams, member progress, failures, and overall plan completion in real time — and can send hints to running delegates.

Memory that persists

Babo agents don't forget between sessions:

• Cryptex — layered long-term memory (identity, user model, project facts, skills, channels, and more)
• Working memory — salience-weighted short-term slots for goals, constraints, and active context
• Knowledge & facts — searchable domain memory with conflict resolution
• Merkle chain — tamper-evident conversation and learning history
• Episodes & narrative self — the agent builds a coherent story of your relationship over time
• Sleep cycles — background consolidation turns experiences into durable memory

Explore everything in the Memory view: overview, knowledge tree, chain, working memory, episodes, and soul.

Brain dashboard

The Brain page shows what your agent is "feeling" and doing internally:

• Hormone and drive dynamics
• Signal history (learning, bonding, reflection, and more)
• Working memory and narrative status
• Theory of mind — the agent's model of you
• Network activity over time

It's the control room for a biologically-inspired runtime — useful, transparent, and optional.

Skills & one-click integrations

Install and configure capabilities from the Tools page:

Bundled skills include guided onboarding (QR pairing, conversational setup, or in-app OAuth). Connected MCP tools appear as first-class agent tools in the loop.

Agents can also search ClawHub, install community skills, and crystallize frequently used skills into native plugins.

Built-in agent tools

Beyond integrations, every agent can use:

• Browser automation (navigate, click, snapshot, scrape)
• Shell & files — read, write, search the workspace
• Semantic codebase search
• Scheduler & HTTP pollers — cron jobs and URL monitors
• Contacts — cross-channel address book (WhatsApp, Telegram, email)
• Vision — describe and ask questions about images
• Downloads — offer files back to you in chat

Desktop & self-hosted

• Desktop app (Electron) — bundles the UI + local Python runtime; first-run wizard sets up Python, inference, and backend
• Web UI (Angular) — full product in the browser
• Backend (NestJS + PostgreSQL) — accounts, agents, WebSocket relay, ClawHub proxy
• Local runtime (Python) — agent brain, memory, tools, and FastAPI server on 127.0.0.1:9222

Run desktop-only, self-host everything, or mix cloud auth with a local runtime.

---

Quick start

Desktop (recommended)

Download the latest release (Windows / macOS installers from CI).

Or build from source:

git clone https://github.com/umbecanessa/babo.git
cd babo/desktop
npm install
.\build-local.ps1          # Windows: unpacked app in desktop/release-build/build-*/win-unpacked/
# or: npm run build && npm run dist:win   # NSIS installer
# macOS: npm run dist:mac (see desktop/BUILD-MAC.md)

On first launch: set up Python, your inference URL + model (+ optional API key), and backend URL.

Self-hosted stack

1. Database + backend

docker compose up -d postgres
cd backend
cp .env.example .env   # edit secrets + RUNTIME_URL
npm install
npx prisma migrate deploy
npm run start:dev

2. Agent runtime

pip install -r requirements-desktop.txt
export NLS_VLLM_BASE_URL=https://openrouter.ai/api/v1
export NLS_HF_MODEL=openai/gpt-4o-mini
export NLS_INFERENCE_API_KEY=your-key   # if needed
uvicorn server.main:app --host 127.0.0.1 --port 9222

3. Frontend

cd frontend
npm install
npm start

Open the app, register, create an agent, and start chatting.

See the documentation for full guides.

---

Repository layout

Documentation

• Documentation — MkDocs site at babo.agency
• Documentation home — full map (architecture, reference, extension)
• Deployment topologies — desktop hub + web relay
• Quickstart
• Extension guide

The marketing homepage in website/ is deployed with the docs — GitHub Actions overlays it onto the MkDocs build at babo.agency. See website/README.md and .github/workflows/docs.yml.

---

License

MIT — see LICENSE.