Control a swarm of coding agents by voice, from any browser.
Website · Setup · Security · Mods · Contributing
Seti is a self-hosted control room for Claude Code. It runs your agents on your own Linux machine, each in a sandbox, and puts all of them in one web app you can open from a laptop or a phone.
Every screen answers one question first: what needs you? Agents that wait on a permission or ask a question go to the top. You answer them in a tap, or call the Executive Assistant (EA) and ask out loud. The EA knows what each agent is doing, briefs you on what changed while you were away, and holds routine news until you ask for it.
The agents keep shipping while you're away. Their pull requests come back to them with CI results and review threads, so a red check or a new comment becomes their next turn, not your next chore.
- Voice first. Call the EA from any browser, wake it with "Hey Seti", and hear answers with live captions.
- Needs you, first. One list of the agents that cannot go on without you, each with its reason.
- Full control on a phone. Permissions, terminals with a key row, files and every setting.
- A live graph of the team. Home links every agent to its repos and subagents.
- Pull requests that close the loop. CI, previews and reviews reach the agent as input.
- Channels. Webhooks, reminders and pollers turn any service into events that wake an agent or the EA.
- One group chat. You, the EA and every agent, where an @mention reaches an agent mid-turn.
- Ask aside.
/btwasks an agent a question from its context without interrupting its turn, and the answer stays in its transcript as an aside. - Requests in one place. Agents' questions, access requests, plans, tool calls and knowledge notes wait in the Requests section, where you answer each in place and one Submit all answers every agent's questions at once, or in a drawer beside the agent.
- Shared knowledge. Agents propose lessons for each other, and only the notes you approve reach them.
- Survives restarts. An agent cut off by an upgrade or a reboot picks up where it stopped.
- Mods. Turn features on and off, or build your own, sandboxed with only the access you grant.
Every flow works on a phone, and the EA can do almost everything you can, except what stays yours.
| What you do | Phone | Desktop | Voice (EA) |
|---|---|---|---|
| Start, stop and message agents | Yes | Yes | Yes |
| Answer an agent's question | Yes | Yes | Yes |
| Allow or deny a permission | Yes | Yes | No |
| See pull requests, CI and reviews | Yes | Yes | Yes |
| Post in the group chat | Yes | Yes | Yes |
| Hear what changed while you were away | Yes | Yes | Yes |
| Open a terminal | Yes | Yes | No |
| Read and edit files, share a frozen copy | Yes | Yes | No |
| Send files and links from other apps | Android | No | No |
| Change security settings and secrets | Yes | Yes | No |
![]() Needs you first |
![]() Allow from the phone |
![]() The EA answers aloud |
![]() Terminal with a key row |
![]() Home |
![]() A call with the EA |
||
On Linux with systemd (Ubuntu 24.04 is the reference), as a user with sudo:
curl -fsSL https://raw.githubusercontent.com/rickaym/seti/main/scripts/install.sh | bash -s -- --with-eaThe installer asks before each step that needs sudo, sets up PostgreSQL, Node.js, bubblewrap, gh
and the Claude Code CLI, and starts Seti as a systemd user service on 127.0.0.1:3006. Leave out
--with-ea to skip the voice assistant, or add --with-redis for Seti's own cache. You can read
the install script before you run it.
- Open a new shell, then sign in to Claude Code (
claude, then/login) and to GitHub (gh auth login). - Open
http://seti.localhost:3006. From another machine, forward the port first:ssh -L 3006:127.0.0.1:3006 <host>. - Sign in with the token that
cat ~/.seti/credentials/local-token-userprints. - Start an agent: say what it should do and which repos to clone.
Setup covers every option, remote access and upgrades. pnpm run doctor in
~/seti says what is wrong and how to fix it.
- One server owns every session. A Node server keeps one Claude Code process open per agent, so follow-ups, interrupts and PR updates reach the same session.
- Each agent runs in a sandbox. bubblewrap gives it its own PID and network namespaces and only its own folders, and a per-agent proxy is its only way out.
- Keys stay outside. Agents hold placeholders; the proxy adds the real token on allowed hosts only.
- The EA talks to the same API. A Python voice and chat service acts through Seti's API, and the server refuses it anything that touches security or secrets.
- Everything is recorded. PostgreSQL holds the state, and every action lands in an audit log kept as a hash chain.
The SETI Protocol specifies this behaviour as numbered requirements that another implementation can follow, and Implementation maps each part of it to the code.
Agents run code on your machine, so Seti treats every agent as untrusted.
- Sandboxed. Agents see only their own folders and reach only the hosts you list.
- No keys inside. GitHub tokens, cloud roles and API keys are added by the proxy on the way out.
- Guards that always hold.
rm -routside the agent's folders, killing by name and writes to credential folders are refused, even in Bypass, which is per agent, off by default and expires. - Yours alone. The EA and agents cannot change security settings, read secrets or approve shared knowledge. An agent that needs a key asks, and you paste it on Allow.
- Closed by default. Seti listens on loopback, and every call needs a passkey, the local token or a Cloudflare Access login.
- Shares are copies. A share link serves a frozen copy with no script and no route to the file system.
Security has the trust model, hardening steps, where each secret lives, and rotation. To report a vulnerability, see SECURITY.md.
Does Seti run on macOS or Windows? The server needs Linux with systemd, because the sandbox is built on bubblewrap. You can use it from any browser on any device.
Which agents does it run? Claude Code, through the Claude Agent SDK, signed in with your own Claude account.
Do I need the voice assistant? No. Seti works without it. Chat with the EA needs an OpenAI key, and voice also needs Deepgram and ElevenLabs keys.
Is there a phone app? Seti is a web app that works in a phone browser, and you can add it to the Home Screen. iPhone and iPad shows how.
How do I reach it away from home? Put it behind Cloudflare Access, or forward the port over SSH. Setup covers both.
Where does my code go? It stays on your machine. Agents send what they read to Anthropic as Claude Code does, and the EA sends what you say to the voice providers you set up.
Issues and pull requests are welcome. CONTRIBUTING.md has the checks, the conventions and the security checklist. Mods and Channels are good places to start.
Apache License 2.0. The UI shell is adapted from the Homer 11 UI; the favicon is the
Full moon icon from Flaticon; the
Catppuccin palette (MIT) and
JetBrains Mono (SIL OFL 1.1) are used under their licenses. The
Twemoji mod draws emoji with Twemoji graphics, copyright
Twitter, Inc and other contributors, licensed under
CC-BY 4.0; mods/twemoji/LICENSE has the details.





