yothere

Run five agents. Watch none of them. yothere advances up to five agent turns at once on the Claude or ChatGPT subscription you already pay for, and gives you one inbox instead of five terminal tabs.

The hosted cockpit lives at app.yothere.ai (invite-only beta). It holds your account, your threads, and the UI. The agents run on your own machine, on your own Claude Code or Codex CLI login. yothere never holds your model credentials and never bills you for inference.

Product docs (canonical, kept in step with what ships): yothere.ai/docs. This README is the repo front door.

What you need

• A Mac or a Linux box you can leave open. Those are the only two platforms the service installer supports; on Windows, run it inside WSL2.
• Python 3.11, 3.12 or 3.13. 3.14 is not supported and yothere doctor fails it.
• pipx (brew install pipx on macOS).
• Your own coding agent, logged in: Claude Code (the claude CLI on PATH) or the Codex CLI on a ChatGPT subscription.

And you leave the laptop open. This is the consequence of BYO compute, so it is worth saying plainly: while your paired machine is asleep or offline, nothing advances. Dispatched tasks queue and drain the moment it reconnects. Nothing is lost. Nothing moves either.

Hosted beta, end to end

The full walkthrough is docs/BETA-GUIDE.md. The short version:

# 0. Redeem your invite code at https://app.yothere.ai and open the cockpit.
pipx install yothere                                        # Python 3.11-3.13

# 1. Pair this machine (the cockpit's "Connect a machine" panel mints the token).
yothere login --url https://app.yothere.ai --token <paste>

# 2. Start the leaser in the FOREGROUND once. It prints the unattended-tool-use
#    consent and refuses to start until you acknowledge it. Read it before you ack.
cd <the project you want the agent to work in>
yothere service

# 3. Once you have acknowledged it, make it always-on (survives logout and reboot).
yothere service install
yothere service status      # expect: running

# Anytime:
yothere devices list        # paired machines (yothere devices revoke <id> to kill one)
yothere doctor              # diagnose the install (--bundle writes a redacted tarball)

Then dispatch from the cockpit. For voice, click Connect in the cockpit: voice is hosted, so there is no phone number, no Twilio, and no API key of your own.

Read this before yothere service install

Worker turns run your agent with permissions bypassed: claude -p --permission-mode bypassPermissions, or codex exec -s danger-full-access. Inside a turn the agent reads, writes and runs commands in YOTHERE_WORK_DIR without asking you per action.

• yothere service (foreground) gates on that: it prints the consent text and blocks until you set YOTHERE_BYPASS_PERMISSIONS_ACK=1.
• yothere service install does not print it. The installed unit runs python -m yothere.cloud.agent directly. Run the foreground path once first, so the ack is recorded in ~/.yothere/yothere.env before the always-on unit ever starts.
• Point YOTHERE_WORK_DIR at a scratch or project directory, never at something you would mind an agent editing.
• Want the unattended bypass off entirely: set YOTHERE_WORKER_PERMISSION_MODE=acceptEdits in ~/.yothere/yothere.env.

What yothere does contain is outward actions. A send-gate on your machine denies sends, posts and pushes in an unattended turn: the thread stops at blocked and waits for your approval. git push and gh are in a never-grant list, so no approval can ever unlock them for a worker turn: an agent can never publish. You still can, and yothere carries it out: press Ship it on a diff and it pushes the branch and opens the PR as you (publish.py, host-side, outside the turn, plain git with no model in the loop). Cost caps run per thread and per day.

The gate is a guardrail, not a sandbox. It inspects shell and MCP tool calls, so what an agent reads, and its outbound GET requests, are not contained. See SECURITY.md.

Install

pipx install yothere                  # macOS: pip is usually not on PATH, and PEP 668
                                      # blocks a global install. pipx is the path.
pipx install 'yothere[mcp]'           # + the MCP surface (drive a fleet from any MCP client)

State roots at ~/.yothere (override with YOTHERE_HOME; an existing ~/.relay is used as a fallback). YOTHERE_* is the canonical env namespace; legacy RELAY_* names still resolve as deprecated aliases. Every seam: docs/configuration.md.

Commands

yothere spawn "<task>"        # create a thread (or: yothere preset <key> <arg>)
yothere workspace add <name> <path/to/repo>   # register a repo you already have
yothere spawn --repo <name> "<task>"          # a code thread, in its own git worktree
yothere list                  # live threads + state
yothere show <id>             # status, plan, transcript
yothere reply <id> "<text>"   # answer a blocked thread -> it re-advances
yothere focus <id>            # pin the focus thread (yothere unfocus clears it)
yothere board --open          # the glance board
yothere note "<text>"         # capture a reminder (no thread)
yothere park <id>             # pause a thread (yothere kill <id> removes it)
yothere tasks                 # brain-discovered task source (discover|refresh|show)
yothere feedback "<text>"     # send feedback to the team (paired) or save it locally
yothere doctor                # pass/warn/fail per check, with the fix

Full reference: yothere.ai/docs/cli.

The code lane

Register a repo you already have on the machine. yothere never clones, and never holds a git credential of yours:

yothere workspace add site ~/code/site        # also: workspace list, workspace rm <name>
yothere spawn --repo site "fix the layout shift on the pricing page"

The thread's turns run in their own git worktree (under ~/.yothere/worktrees/, not in your checkout), on their own branch. So a code thread cannot tread on what you have open, or on another thread. It commits, and hands the branch back for you to read.

No approval opens a push, a PR or a merge. git push and gh are in the never-grant list, so no approval unlocks them, not yours and not ours. You publish, or nobody does. That is a rule in the gate, not a sandbox: YOTHERE_WORKER_SEND_POLICY=open turns the gate off wholesale, and a GitHub MCP server you install yourself can carry a publish tool the gate does not read as a send.

The worktree is isolation from other threads, not a sandbox: a bypassed turn can still write elsewhere on the box. See SECURITY.md.

Coming soon

One item, marked here so nobody plans around it:

• Recurring tasks. There is no scheduler. A hail runs once, immediately. "Every morning at 7, scan the news" spawns one thread that starts now.

Answered in the FAQ: yothere.ai/docs/faq. The whole picture: yothere.ai/roadmap.

Self-host, and your own brain

Everything above is the hosted product. The repo also runs standalone.

A brain is whatever actually does the work. yothere talks to it over one published wire contract, docs/brain-protocol-v1.md (WebSocket + JSON-RPC 2.0): hello / streamSubscribe / prompt / cancel / close, streaming back delta, and optionally progress, status and cost. The reference implementation is src/yothere/voicecall/echo_brain.py, about 60 lines, and it is the conformance fixture.

The harnesses that run today: claude (Claude Code, local), codex (Codex CLI, local, on a ChatGPT subscription), and remote (anything speaking the Brain Protocol).

# Drive the bundled reference brain, no keys, no account.
python -m yothere.voicecall.echo_brain --port 9999 &
export YOTHERE_REMOTE_BRAIN_URL=ws://127.0.0.1:9999
export YOTHERE_THREAD_HARNESS=remote
yothere spawn "research agentic commerce"
python -m yothere.runner once     # advance one tick (or: python -m yothere.runner loop)
yothere board --open

Two load-bearing caveats live in SECURITY.md: a remote brain's cost cap is advisory, and the brain owns its own safety and permissions. Local walkthrough: docs/ONBOARDING.md and QUICKSTART.md.

Architecture

            yothere (cockpit + voice + runner)            YOUR BRAIN
  ┌───────────────────────────────────────────┐   ┌──────────────────┐
  cli  ──► spawn ──► store (dir-per-thread) ◄── runner ─┐            │
   │         │            │                       │     │  Brain     │
  board ◄─ fleet_state ◄──┘                  brain_advance ──ws──►  Protocol  │
   │     (attention router: rank · focus)          │     │  v1        │
  voice ──► session_manager ──────────────────────┘     └──────────────────┘

• store / thread_model. The dir-per-thread state machine (atomic status.json, session-id resume, the worker contract).
• runner / worker. The headless advance engine. Up to 5 work turns in parallel (YOTHERE_MAX_CONCURRENCY), plus a separate pool for typed questions (YOTHERE_CHAT_CONCURRENCY), so your reply never queues behind the work. Cost caps, stale sweep, 429 usage-cap hold.
• attention. Deterministic ranking of what needs a human.
• brain/. The harness clients: local Claude, local Codex, any remote brain.
• cloud/. The hosted control plane (auth, jobs queue, voice worker) and the BYO-compute leaser that long-polls it.
• board / card. The glance UI (server-rendered, XSS-safe).

Hosted mode (multi-tenant)

YOTHERE_AUTH_MODE selects the cockpit's mode:

• off (default), local, single-user, gated. Every route except the liveness probe, the PWA statics and /login requires loopback, the opt-in trusted tailnet identity, the YOTHERE_VOICECALL_BEARER bearer, or a signed local-session cookie.
• hosted, multi-tenant. Login required, no tailnet carve-out, and each account is scoped to its own ~/.yothere-<tenant> home, so two logins see two isolated fleets. Auth is stdlib-only (yothere.voicecall.auth): scrypt hashing, opaque session tokens, login lockout, Secure cookie tied to the mode.

This is what runs at app.yothere.ai (Fly + managed Postgres), with invite-gated signup and zero agent compute on the control plane: each tenant pairs their own machine, which long-polls a device-token job queue and runs the turns locally. Every paired machine carries a permission tier (read-only by default, up to full), chosen in the pairing ladder and editable in Settings → Machines; the control plane refuses to lease work to a restricted machine whose wheel cannot enforce its tier. Deploy notes: docs/DEPLOY-fly.md.

MCP surface

Drive a fleet from any MCP client over stdio:

• spawn_thread(task, mode, focus): spawn thread(s) from a natural-language task.
• board(): the fleet at a glance, needs-eyes first.
• reply(thread_id, text): approve, edit or reject a blocked thread.

// Claude Desktop: claude_desktop_config.json
{ "mcpServers": { "yothere": { "command": "yothere-mcp" } } }

The tool logic lives in SDK-free _*_impl helpers (yothere.mcp_server), so it imports and tests without the mcp dependency installed.

Status

Invite-only beta. The installed version is whatever pipx install yothere resolves to; yothere --version prints it, and the current release is on PyPI and in GitHub Releases. What changed in each one: CHANGELOG.md. Beta walkthrough: docs/BETA-GUIDE.md.

Live today: the hosted cockpit and inbox at app.yothere.ai (invite-gated signup), up to five parallel turns per paired machine, the code lane (a thread bound to a repo you already have, working in its own git worktree), hosted voice with a daily cost cap, web push to a phone, the send-gate and cost caps, and the BYO-compute leaser. Not live: scheduling (see Coming soon).

Hosted voice runs a dedicated voice-worker app. The per-tenant voice_media flag still defaults to laptop, so a new account is flipped to the hosted plane by hand until the default moves.

Invites and support: hey@yothere.ai.

Contributing

Two people build yothere and neither should break main alone. Every change goes through a PR, and the gate is green CI rather than the other founder: either of us merges our own PR once the checks pass. See CONTRIBUTING.md.

License

MIT, see LICENSE.