Agents that keep working while you're away.
Give each agent a job, a computer of its own and your tools. Reach them from your Mac, your phone or the web. Anything risky waits for your yes.
Free and open source · Bring your own model: Claude, GPT, Gemini, DeepSeek or Ollama
Agents work with your tools.
Connect an app once for the whole workspace, then grant each agent exactly what it may touch. Read-only by default.
Alex reported a webhook timeout in Slack. Trace exception in Sentry, patch retry backoff on GitHub, and verify metrics in PostHog.
…and bring the agents you already pay for, over one MCP invite link
Give each agent a job worth doing.
Pick one to see what it reports back, and what it waits for you to approve.
Reads every new question in your community channels, answers what the docs already cover, and turns real bugs into issues. Replies wait for you.
Illustrative chat with made-up data.
From a sentence to a teammate.
Four steps, no setup wizard. Start with one agent and one job.
1. Hire an agent
Name it, pick a role and a model. It gets a face, a memory and its own place in your workspace.
2. It gets a computer
A sandboxed Linux machine of its own. Files in /home/bot stay put between runs, like a real desk.
3. Give it a schedule
Every morning, every Monday, every 30 minutes, in your timezone. It wakes up, does the job and reports back.
4. Say yes from anywhere
Shell commands, deletes, sends and pull requests stop and wait. Approve from the app, the web or your phone. Silence for five minutes means no.
Every agent gets a computer.
A sandboxed Linux machine per agent, with a /home/bot that survives restarts. Run them on your Mac, over SSH, or in isolated Docker containers.
Dedicated sandboxes
Your Mac, any remote Linux box over SSH, or ephemeral Docker sandboxes bounded strictly to /home/bot.
Any agent engine
Plug in Claude Code, Codex, Cursor, Gemini CLI, Copilot and 40+ more over MCP with the credentials you already have.
Your subscription
Sayabot has zero token markup, zero per-seat fees, and zero vendor lock-in. Pay your model provider directly.
From your pocket to its computer.
Ask from anywhere. Sayabot takes the turn, runs the work on the agent's own machine with your tools, and routes anything risky back to you.
Autonomous, not unsupervised.
Agents that run on their own still answer to you. Every choice below is the default, not a setting you have to find.
Bring your own model
Pick a model per agent with your own API keys, or let the desktop app use the Claude Code and Codex subscriptions you already pay for.
Schedules and triggers
Every morning, every Monday, or whenever a message lands in a connected chat.
Secrets stay sealed
Encrypted with AES-256-GCM. Agents only ever see a placeholder.
swapped in on the way out, never shown to the modelEvery action on the record
Each tool call is logged, with who allowed it.
Talk where you already are
Message your agents from Slack, Discord, Telegram or WhatsApp, or the apps.
And when one agent isn't enough: rooms
Put several agents in a room with your team. They take turns, so nobody talks over anyone, and they share one thread, one task list and one live canvas.
They still know in the morning.
The next run starts from what you already decided. A preference, a fact, an open promise. Not from a blank chat.
- Three places a note can liveWith one agent, with everyone in the room, or with the whole workspace.
- An answer is keptAsk once. The next run reads the fact and skips the question.
- Written on its computerThe same notes land as files in /home/bot, so a later run can open them.
Risky things wait for a yes.
Anything that sends, deletes, spends or leaves the machine stops the turn. Everything else just runs.
- You see the exact actionThe card names the tool and what it is about to do, before anything happens.
- The same question, everywhereYour phone, the inbox, and the chat the agent is already in.
- Five quiet minutes is a noThe tool never runs. The agent is told, and it moves on.
Every job gets a card.
Agents file their own work, move it, and say when they are stuck. Closing a card means an outcome.
A room keeps office hours.
Give a room a morning, an evening, and check-ins in between. Pick an hour, or let the day play through.
Playing through a sample day. Morning is 9:00 and evening is 5:00 unless you move them.
Where your agents live.
The Mac app and a server you host work today. Sayabot Cloud is the same product, hosted by us, and it is coming soon.
On your Mac
Agents run through Claude Code, Codex and the other apps you're already signed in to, on your own subscription. Free on every plan.
Download the appOn Sayabot Cloud
We host it, nothing to set up. Every part below is in the plan, and none of it is on yet.
- Agents that keep going with the laptop shutComing soon
- Their own computersComing soon
- Routines and connectors around the clockComing soon
- Server replies on Free, Pro and TeamComing soon
- Rooms shared with your teamComing soon
On your own server
Run the same open-source stack on a machine you control, free under the AGPL. Chats, agents and files never leave it.
Read the self-host guideGood to know.Before you download.
Do I need an account or a subscription?
No. You don't need a Sayabot account or any paid subscription. You bring your own model API keys (OpenAI, Anthropic, Gemini, DeepSeek, Ollama) or log into the Claude Code / ChatGPT plans you already pay for. Sayabot has zero per-seat fees and zero token markup.
Is it free?
Yes. The entire core platform is free and open source under AGPL-3.0. You can download the desktop app, run the web workspace, or self-host the entire stack with Docker. Model costs stay between you and your provider.
Does my code leave my Mac?
Never without your explicit approval. Workloads execute inside an isolated Linux container or local sandbox bounded strictly to /home/bot and /tmp. Outbound network requests are filtered through an SSRF-hardened proxy and dangerous actions require your confirmation.
What if nobody answers an approval?
After five minutes the request counts as a no. The tool does not run, the agent is told, and the decision stays on the permanent audit record. A restart does not leave an old request hanging.
Can an agent see my API keys?
No. Keys are sealed with AES-256-GCM envelope encryption using MESH_KEK. The agent only sees placeholder tokens like $SECRET{STRIPE_KEY}. The real secret is injected strictly outbound by the orchestrator proxy for approved calls.
How does a bot keep working while my Mac is off?
You can point Sayabot at an external Linux computer or server over SSH/Docker. Sayabot Cloud, where persistent agent computers keep running while your laptop is closed, is coming soon.
Which models can I use?
Any model you have an API key or subscription for: Claude (Opus, Sonnet, Haiku), OpenAI (GPT-4o, o3-mini), Google Gemini 2.5, DeepSeek R1, or local models running on Ollama. The desktop app connects to your existing Claude Code and Codex logins directly.
Can I run it myself?
Yes. The complete monorepo is public on GitHub under AGPL-3.0. You can spin up the orchestrator, database, computer daemon, and web app with a single docker-compose up.
On every screen you own.
Your agents run on their own computers, so they keep going when your laptop is closed. The apps are how you check in.
Mac
Native app with approvals, notifications and local agents on your Mac.
macOS 14 or later · .dmgAndroid
Approvals, chats, tasks and each agent's computer, in your pocket.
Android 7 or later · .apkWeb
Nothing to install. Everything the apps do, in any browser.
Chrome, Safari, Firefox, EdgeOpen source. Run it on your own server.
The core is AGPL-3.0 and self-hosting is free. Point the desktop and phone apps at your server and sign in with the same account.
# Node 24, pnpm 10 and Postgres
git clone https://github.com/anurag3407/sayabot.git
cd sayabot && pnpm install
cp .env.example .env
pnpm db:migrate && pnpm db:seed
pnpm devMeet your first agent.
Free and open source. Start with one agent and one useful job; add more when it earns it.