Ciaobot

Ciaobot is a second brain you own — a local, provider-agnostic AI workspace whose memory is a plain-markdown vault you own outright. Chats, projects, files, schedules, and memory live in one interface, backed by whichever model you choose (Claude Code, OpenAI Codex, and others).

Install

macOS 13+ (Homebrew) — recommended; installs Ciaobot.app plus the engine:

brew install --cask raffaelefarinaro/ciaobot/ciaobot-desktop

The app is ad-hoc signed and not notarized. If macOS blocks the first launch, Control-click Ciaobot.app, choose Open, and confirm once. Do not disable Gatekeeper. First launch opens the setup wizard when no server is configured.

Already using the Homebrew engine from an earlier release? Move to the app without recreating your workspace:

brew update
brew upgrade ciaobot
brew install --cask raffaelefarinaro/ciaobot/ciaobot-desktop
open -a Ciaobot

The first app launch reuses the existing workspace and server LaunchAgent, disables the legacy menu-bar helper, and moves the old Ciaobot Server.app to the Trash once the engine is reachable. Browser-installed PWA shortcuts are left alone. Future releases use the app's single Update… action to update the engine and desktop app together.

Any platform (PyPI) — or macOS without Homebrew; requires Python 3.12 or newer:

python3.13 -m venv ~/.ciaobot-venv
~/.ciaobot-venv/bin/pip install ciaobot
~/.ciaobot-venv/bin/ciao run

Then open http://localhost:8443 and follow the setup wizard:

• Workspace folder (default ~/ciaobot) — your second brain (memory-vault/) plus app config and runtime state. Sync this folder (GitHub, Drive, iCloud, …) so your vault follows you across machines.
• Model provider — Claude Code, Codex, or another configured backend.

The wizard writes config, initializes the workspace as a git repo (with a .gitignore for secrets and runtime state), and installs the macOS engine LaunchAgent. A cask installation uses Ciaobot.app; package-only installs retain the legacy recovery launcher during the migration release.

For scripted setups: ciao setup --workspace <dir>. If a setup link returns invalid setup token, mint a fresh one with ciao setup-url --workspace <dir>.

Contributors running from a git checkout: see docs/DEVELOPMENT.md.

Why it exists

Ciaobot started from the limits of Claude Cowork. Cowork is genuinely good at doing the work — it runs on your own desktop, edits your own files, and follows along on your phone. But it's Claude-only, and everything it learns about you — your sessions, your context, your history — lives in an account you rent. It knows you no better on day 365 than on day 1: it remembers sessions, not you.

Ciaobot is the other bet — differences that are structural, not features:

• Any model, local or cloud. Claude Code, OpenAI Codex, Ollama (local or cloud), or OpenRouter — the second brain outlives any single model, and lighter tasks can run entirely on your own hardware.
• Your machine's own power. Voice is already wired to work on-device: dictation and read-aloud run on your computer (Apple Silicon) with no cloud round-trip needed.
• Accrual, not storage. A memory system, baked in, that curates itself and compounds into you over time — not a pile of past sessions.
• Yours to keep. That memory is plain markdown in a git repo you own outright — portable, and useful even if this app disappears.

The coworker is just the interface. The second brain is the product.

It follows you — without leaving your machine

Ciaobot runs on your computer, but you're not chained to your desk. The PWA is built mobile-first with push notifications, and over Tailscale your phone reaches the machine at home on a private mesh — the same desk-to-phone continuity as a cloud assistant, without shipping your work to anyone's servers. The one honest trade: the host has to be awake. When the machine sleeps, so does your second brain.

Who it's for

Ciaobot is built for knowledge work, not software development: brainstorming, research, writing and editing, planning, and document work — typically drafted as markdown in a local vault, then published to Google (Docs, Drive, Sheets) when ready.

• Not built for day-to-day coding. There is no code editor or repo tooling in the UI — keep using your IDE for that.
• Google Workspace — Gmail, Calendar, Drive, Docs, Sheets, Slides, and Tasks through Google's gws CLI, connected with browser-based OAuth from Settings.

The idea

Ciaobot does not reinvent how you talk to agents. It runs Claude Code or OpenAI Codex in the background, on the bet that the vendors' own CLIs are the best-maintained agent harnesses available — they keep the model communication, tool use, and agentic loop optimized so this project doesn't have to. Ciaobot stays in control of the three things that matter to me:

1. The context — deciding exactly what memory, notes, and project state the agent is fed each turn.
2. One interface — the same UI regardless of which project or provider you're talking to.
3. Incremental capabilities — features are added only when I need them or discover a pattern worth adopting, not speculatively.

The same instinct runs through the infrastructure: rather than build and run servers, Ciaobot stands on primitives that already have millions of users hardening them — git for sync, versioning, and durability; Tailscale for private remote access; and one PWA shared by browsers, phones, and the thin macOS shell. Every hard problem is answered by borrowing a battle-tested tool, not maintaining a new one.

What that looks like in practice:

• Workspaces and projects — split life areas (personal, work, a client, …) into sidebar workspaces, then organize work inside projects. Ciaobot injects project notes and context into every turn.
• A vault you own — durable knowledge as plain markdown with wikilinks and an INDEX.md, inspired by Andrej Karpathy's LLM Wiki pattern. Browse in Obsidian or any editor; sync via GitHub, Drive, or iCloud.
• Skills, subagents, and commands — packaged defaults, extensible from Settings or workspace files (see What ships by default).
• Files and automations — create, preview, edit, and restore vault files from the UI; run recurring routines on a cron you choose (schedules) or re-run a prompt inside one chat every N minutes (loops).
• Voice, notifications, and updates — transcription, push alerts, model settings, and in-app package updates. On macOS, Ciaobot.app owns the window, menu bar, native notifications, and desktop updates while the engine runs as a background service.
• Provider choice — Claude Code or Codex with your existing login; Ollama, OpenRouter, and on-device models for lighter tasks (see Providers).
• Agent-safe control plane — an authenticated, chat-scoped MCP surface lets managed Claude Code and Codex processes operate Ciaobot memory, vault, projects, chats, schedules, loops, agent handoffs, and file history without curl or direct runtime-JSON edits. MCP is the default transport, with the legacy CLI path retained as an automatic fallback. Reads and non-destructive writes on this surface run without an approval card (they are the twins of buttons in the UI); deletes and lifecycle actions still ask. See docs/MCP.md.

Pick a workspace folder, choose a provider, and work — Ciaobot is the interface on top; the vault is yours to keep.

Memory and the vault

Ciaobot keeps memory in layers so the agent can recall what matters without stuffing every prompt. Settings → Context shows what the agent actually loads.

• Short agent memory (~/.ciao/memory.md and user.md) — a small, capped scratchpad the model maintains for you: preferences, conventions, lessons. Updated during conversation or via /remember; a snapshot is injected at the start of each chat. Add [expires: YYYY-MM-DD] to a temporary entry to keep it active through that date. It is hidden from later snapshots, but still uses stored character budget until daily memory curation removes it.
• Your vault (memory-vault/, or a separate vault root per sidebar workspace) — durable markdown you own: people, projects, ideas. Browse it in Obsidian or any editor; it stays useful even without Ciaobot.
• One behavior file for the install — <workspace>/CLAUDE.md (and AGENTS.md for Codex) applies to every chat.

When your message mentions a name that appears in the vault index, the agent gets a quiet hint — “this probably means People/Emma” — so it opens the right note without you repeating context. And when a chat is archived, a pipeline turns it into durable knowledge: session insights are extracted, memory proposals are drafted, and daily/weekly curation runs update vault pages — but nothing is promoted into long-term memory without review, and Ciaobot never discards or rewrites an existing notes folder during onboarding. Track the background steps under Settings → Automation, and see docs/ARCHITECTURE.md for the full pipeline.

Features

Chats, projects, and workspaces

• Sidebar workspaces per life area (personal, work, a client) — each with its own vault, projects, and default model.
• Projects group related chats and inject durable notes and context into every turn.
• Comment on any passage of a reply — select text, attach a note, and it rides along with your next prompt; queue follow-ups while the agent is still working.
• Per-chat model picker with provider thinking levels on top of per-workspace defaults.
• Fork conversation: create a new independent chat in the same project starting from any completed agent answer, preserving history.
• Agent handoffs: spawn and communicate with a second provider route (the participant) as a read-only sub-chat attached to the originating turn.

Voice — dictation and read-aloud

• Speech-to-text dictation in any chat: cloud transcription or free on-device via mlx-whisper (Apple Silicon).
• Text-to-speech read-aloud of replies: cloud voices or free on-device via Kokoro; local models download on first use and are re-installed automatically after app upgrades.

Files and documents

• Completed turns surface touched files as clickable Outputs chips below the final reply. Expand Activity to inspect the chronological notes, tool calls, and file cards; click a file for history, diff, and restore.
• Pin a document beside the chat and add line-level comments on the preview (attached to your next message, like chat comments).
• Rich previews: images inline, PDFs in a built-in viewer, PowerPoint (.pptx) converted to PDF for display (requires LibreOffice on the machine running Ciaobot).
• Create, edit, and restore vault files from the UI, with snapshots behind every agent edit.

Memory, vault, and insight extraction

• Layered memory: a capped agent memory and user profile injected at chat start, plus a plain-markdown vault you own (Obsidian-compatible, syncable via GitHub/Drive/iCloud).
• Archiving a chat runs an extraction pipeline: session insights and trajectories are captured, memory proposals are drafted, and project canonical docs are updated — nothing is promoted into long-term memory without review.
• Daily and weekly curation routines keep vault pages and Workspace/Learnings.md current.
• Vault-index hints: mention a name the index knows and the agent is quietly pointed at the right note.

Automations

• Schedules: recurring or one-off cron routines that dispatch fresh prompts into a project or chat.
• Loops: re-run a prompt inside one chat every N minutes, keeping the conversation's context between iterations. A loop-driven chat is marked with ↻ in the sidebar and on the home grid, and carries a banner with start/stop controls; the harness's own /schedule and /loop skills are removed from the model's context and denied, so automations land in Ciaobot instead of a cloud routine it cannot see.
• System routines ship enabled (memory curation, workspace hygiene, skill evolution); every background run is visible under Settings → Automation.

Extensibility — skills, subagents, commands

• Stock skills, subagents, and slash commands ship with the app; same-named workspace versions override them.
• Install skills from GitHub repositories; they refresh automatically on restart.
• A weekly skill-evolution routine proposes improvements from real usage — reviewable proposals, never silent edits.

Providers and models

• Claude Code or OpenAI Codex with the subscription login you already have; Ollama (cloud or local) and OpenRouter as API backends.
• haiku/sonnet/opus tier routing mapped across providers; background tasks (titles, insights) routable to cheaper or on-device models (apfel).

Google Workspace

• Gmail, Calendar, Drive, Docs, Sheets, Slides, and Tasks through Google's gws CLI, connected with browser OAuth from Settings — no terminal required.

App surface

• Installable PWA with web-push notifications and in-app package updates.
• macOS desktop app: one Dock window and menu-bar companion with native notifications, updates, and a launchd-managed engine.
• First-run product tour plus a getting-started checklist whose steps deep-link into the real pages.
• A local HTTP API an in-chat agent can drive (create chats, subagents, commands — see PWA_API.md).

On first launch, an in-app product tour walks through the core flows. Replay it anytime from Settings → Home → Product tour.

What ships by default

Every install seeds a set of subagents, slash commands, and system routines from the package (ciao/stock/); your own workspace versions with the same name take precedence.

Subagents

Specialized roles the main agent can delegate to (ciao/stock/agents/):

Slash commands

Type these in any chat (ciao/stock/commands/):

System routines

Recurring schedules that ship enabled (ciao/stock/schedules.json); they run through the same provider pipeline as a chat turn, and their runs are visible under Settings → Automation:

Your own schedules live alongside these in the workspace (.runtime/schedules.json), with in-chat loops in .runtime/loops.json; both are managed from the UI's Automations page. Packaged skills (vault search, Google Workspace, web research, and more) are browsable under Settings → Assets and live in ciao/stock/skills/.

Providers

Use the access you already have:

• Claude Code — CLI-managed Claude subscription or Anthropic Console authentication.
• OpenAI Codex — codex login, including eligible ChatGPT subscription accounts.
• Ollama — cloud or local daemon.
• OpenRouter — OPENROUTER_API_KEY.
• On-device models — for lightweight tasks where available: titles via apfel, speech via mlx-whisper, and similar.

See INTEGRATIONS.md for env vars, OAuth, and per-task model routing (titles, insights, voice).

A personal project, shared

Ciaobot is my idea of an AI assistant that belongs to you, not to a vendor: it runs on your machine, talks to whichever model you choose, and turns your work into a second brain you keep in plain files. I built it for my own use and run it on my own machines; the defaults reflect that: project-first navigation, a plain-markdown vault as memory, explicit model routing, and self-improvement loops that propose changes instead of applying them blindly.

I'm sharing it because the patterns may be useful to you. Ideas, bug reports, disagreements with my defaults, and pull requests (#agentswelcome) are welcome — see CONTRIBUTING.md.

Documentation

Naming note: the user-facing product is Ciaobot. The CLI is installed as both ciaobot and ciao (same command); the Python package, import path, and many environment variables are still named ciao/CIAO_* for compatibility.

Why "Ciao"?

Ciao isn't just Italian for "hi" and "bye" — it comes from the Venetian phrase s-ciào vostro ("[I am] your slave"), a servile greeting that shed its literal meaning over the centuries and became the everyday word Italians use today. Fitting for an assistant: yours to command. See the etymology on Wikipedia.

Built on

Ciaobot is glue around a lot of excellent open tools — Claude Code, the Claude Agent SDK, Codex CLI, Starlette, Vue, and more. See docs/CREDITS.md for the full list.