Limbo is a second brain with a conversational interface. It stores atomic notes in a local vault, searches them semantically, and maintains Maps of Content (MOCs) to keep knowledge navigable. Runs in a Docker container, accessible via Telegram or the OpenClaw gateway.

Install

Limbo is designed to run on a VPS (always-on, accessible from anywhere). A $5/month Ubuntu server is all you need.

1. Provision a server

Any Ubuntu/Debian VPS with 1 GB+ RAM.

2. Run the installer

curl -fsSL https://get.heylimbo.com/install.sh | bash

This installs Docker, Node.js, and the Limbo CLI.

3. Start Limbo

limbo start

The setup wizard walks you through:

• [ ] Choose a language (English / Español)
• [ ] Select a provider (Anthropic, OpenAI, OpenRouter)
• [ ] Authenticate (API key or Claude/ChatGPT subscription)
• [ ] Pick a model
• [ ] Connect Telegram (optional but recommended)
• [ ] Enable voice messages and web search (optional)
• [ ] Review and confirm

Once complete, Limbo restarts and is ready to use.

4. Update

limbo update

Pulls the latest image and restarts. Vault data is persisted and not affected.

Local Install (macOS/Linux)

If you prefer running locally instead of a VPS:

npx limbo-ai start

Requires Docker Desktop and Node.js 18+. Binds to 127.0.0.1:18789.

Commands

limbo start                  # Install and start (enters wizard on first run)
limbo stop                   # Stop the container
limbo update                 # Pull latest image and restart
limbo status                 # Show container status
limbo logs                   # Tail container logs
limbo start --reconfigure    # Re-run the setup wizard
limbo config voice --enable --api-key gsk_xxx   # Enable voice transcription
limbo config web-search --enable --api-key BSA_xxx  # Enable web search

Connecting

Telegram (recommended)

The setup wizard walks you through creating a Telegram bot and pairing it. Message your bot and Limbo responds — full agent with personality, memory logic, and vault tools.

OpenClaw gateway

Any OpenClaw-compatible client can connect via WebSocket:

ws://localhost:18789

MCP (for other AI agents)

Add Limbo as an MCP server to give another agent direct vault access:

{
  "mcpServers": {
    "limbo": {
      "url": "ws://localhost:18789"
    }
  }
}

This exposes 4 vault tools (vault_search, vault_read, vault_write_note, vault_update_map). The connecting agent operates on the vault directly — Limbo's LLM is not involved.

Optional Features

Enable during the setup wizard or anytime via CLI.

Voice Messages

Transcribe Telegram voice notes using Groq Whisper.

limbo config voice --enable --api-key gsk_xxx
limbo config voice --disable

Web Search

Real-time web search via Brave Search API.

limbo config web-search --enable --api-key BSAxxx
limbo config web-search --disable

Hardware Requirements

| Tier | RAM | vCPU | Disk | |------|-----|------|------| | Minimum | 512 MB | 1 | 1 GB | | Recommended | 1 GB | 1 | 5 GB | | With other services | 2 GB | 1 | 10 GB |

Limbo uses ~150 MB at rest, peaks ~300 MB during agent runs. CPU usage is negligible.

Architecture

┌──────────────────────────────────────────────────┐
│                Docker Container                  │
│                                                  │
│  ┌──────────────┐                                │
│  │  Supervisor  │  (PID 2, under tini)           │
│  └──┬────┬──┬───┘                                │
│     │    │  │                                    │
│     │    │  └── Control plane    :LIMBO_PORT+2   │ ◄── limbo CLI (host)
│     │    │                                      │
│     │    └───── Wizard port      :LIMBO_PORT+1   │ ◄── connect-calendar /
│     │            (on-demand, forked by spawner)  │     switch-brain browser
│     │                                            │
│     ▼                                            │
│  ┌──────────────┐     ┌────────────────────┐     │
│  │  OpenClaw    │◄───►│  LLM provider      │     │
│  │  gateway     │     │  (anthropic/...)   │     │
│  │  :LIMBO_PORT │     └──────────┬─────────┘     │
│  └───┬──────────┘                │               │
│      │               ┌───────────▼──────────┐    │
│      │  Telegram     │  MCP Server          │    │
│      ├── channel     │  (limbo-vault)       │    │
│      │               └───────────┬──────────┘    │
│      │                           ▼               │
│      │                    /data/vault/           │
│      │                    (markdown notes)       │
└──────┴───────────────────────────────────────────┘
       │
       ▼
   Telegram

• Supervisor — container main process, owns OpenClaw + the wizard control plane
• OpenClaw — Node.js runtime (~150 MB) handling connections, LLM routing, Telegram, and MCP tools
• Control plane — tiny HTTP API the limbo CLI uses to open on-demand wizards (connect-calendar, switch-brain) without restarting the container
• Wizard port — a setup-server child spawned on demand, used only during an active wizard
• MCP server — Node.js vault read/write tools, spawned by OpenClaw
• Vault — plain markdown with YAML frontmatter, persisted in a Docker volume

Gateway, wizard, and control plane bind to host loopback only (127.0.0.1). In cloud mode (limbo cloud activate), an additional public server on port 80 makes the wizard accessible via https://{id}.heylimbo.com — no SSH or tunnels needed.

Agent Installation (headless)

For CI/CD or automated provisioning:

npx limbo-ai start --provider anthropic --api-key sk-ant-xxx --model claude-sonnet-4-6

| Flag | Required | Default | Description | |------|----------|---------|-------------| | --provider | yes | — | anthropic, openai, or openrouter | | --api-key | yes | — | Provider API key | | --model | no | Provider default | Model name | | --language | no | en | en or es |

Headless mode skips Telegram. Add it later with limbo start --reconfigure.

Subscription auth (Claude Code, ChatGPT Plus) requires the interactive wizard.

Environment Variables

Managed by limbo start, stored in ~/.limbo/.env.

| Variable | Default | Description | |----------|---------|-------------| | AUTH_MODE | api-key | api-key or subscription | | MODEL_PROVIDER | anthropic | anthropic, openai, openai-codex, or openrouter | | MODEL_NAME | claude-sonnet-4-6 | Model to use | | RUNTIME_REASONING_EFFORT | medium | OpenClaw reasoning effort override | | TELEGRAM_ENABLED | false | Enable Telegram integration | | VOICE_ENABLED | false | Enable Groq voice transcription | | WEB_SEARCH_ENABLED | false | Enable Brave web search |

Development

# Run MCP server locally
cd mcp-server && npm install && VAULT_PATH=./dev-vault node index.js

# Build image locally
docker build -t limbo:dev . && docker compose up -d

# Run tests
npm test

PR evals

Standard PR validation should run from a staging-based worktree, not from your dirty repo checkout and not from a container that silently falls back to latest.

git fetch origin
git worktree add -b codex/pr-242-eval /tmp/limbo-staging-eval-pr242 origin/staging
git -C /tmp/limbo-staging-eval-pr242 merge --no-ff --no-edit origin/<pr-branch>
docker build -t limbo:staging /tmp/limbo-staging-eval-pr242
/tmp/limbo-staging-eval-pr242/evals/scripts/recreate-eval.sh

Notes:

• The eval compose now fails fast unless LIMBO_IMAGE is explicitly set.
• evals/scripts/recreate-eval.sh defaults to limbo:staging and verifies the recreated container is actually using that image.
• Always validate the final container image with docker inspect limbo-eval --format '{{.Config.Image}}' before trusting eval or manual test results.

See CONTRIBUTING.md for release and deployment process.