yothere

An ambient-agent cockpit. yothere is the interface for long-running agents: a fleet board, a voice loop, and a headless runner that advances work and tells a human when to look. The agent that actually does the work, the brain, is anything you point yothere at, over one published wire contract: docs/brain-protocol-v1.md (WebSocket + JSON-RPC 2.0).

yothere is harness-agnostic by design. The same cockpit drives:

• Claude Code running locally (claude -p per thread),
• ubob or any model-agnostic harness over its WebSocket daemon,
• a remote brain over the internet (a hosted sprite, a teammate's stack, a work platform) — anything that speaks the Brain Protocol.

It runs the same whether the brain is on this machine or behind a wss:// endpoint, so you can start a thread at your desk and watch it from your phone.

New here? Start with docs/ONBOARDING.md: a 20-minute walkthrough from install to a real task advancing, plus how to send feedback.

Why it exists

A fleet of agents working in the background is only useful if a human knows which one needs them now. yothere is an attention router for the human, not a work router for the agents: pull-on-glance by default, one rate-limited nudge ("N threads need your eyes"), never a firehose, never interrupting your focus thread.

Install

# macOS: pip is usually not on PATH (only pip3) and PEP 668 blocks global installs,
# so install the CLI with pipx (run `brew install pipx` first if needed).
pipx install yothere                 # core: fleet runner + board + remote-brain client
# Optional extras bake in at install time, e.g.:
#   pipx install 'yothere[voice]'    # + the Gemini-Live voice surface (WebRTC/Twilio)
#   pipx install 'yothere[voice,llm]'# + optional LLM tiebreak for routing (deterministic otherwise)

yothere keeps all of its state under ~/.yothere (override with YOTHERE_HOME, or the legacy RELAY_HOME; an existing ~/.relay is used as a fallback); see docs/configuration.md for every env seam (each RELAY_* name also accepts its YOTHERE_* sibling).

Quickstart — drive the bundled reference brain

# 1. Start the conformance brain (the smallest valid Brain Protocol implementation).
python -m yothere.voicecall.echo_brain --port 9999 &

# 2. Point yothere at it and spawn a thread.
export RELAY_REMOTE_BRAIN_URL=ws://127.0.0.1:9999
export RELAY_THREAD_HARNESS=remote
yothere spawn "research agentic commerce"

# 3. Advance the fleet one tick, then glance at the board.
python -m yothere.runner once   # or: python -m yothere.runner loop  (always-on engine)
yothere board --open

Swap RELAY_REMOTE_BRAIN_URL for your real endpoint and yothere drives your brain. To run threads with a local Claude Code instead, set RELAY_THREAD_HARNESS=claude.

The contract

A brain implements docs/brain-protocol-v1.md: hello / streamSubscribe / prompt / cancel / close, streaming back delta (text), and optionally progress, status (drives the attention router), and cost. The reference is src/yothere/voicecall/echo_brain.py, ~60 lines. Two load-bearing caveats live in SECURITY.md: a remote brain's cost cap is advisory, and the brain owns its own safety/permissions.

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).
• attention — deterministic ranking + the human-facing guardrails.
• runner / worker — the headless advance engine (cost caps, stale sweep, 429 usage-cap hold, coalesced nudges).
• brain/ — the harness clients: local Claude, ubob daemon, any remote brain.
• board / card — the glance UI (server-rendered, XSS-safe).

Hosted mode (multi-tenant cockpit)

The /overview cockpit runs in one of two modes, selected by RELAY_AUTH_MODE:

• off (default) — local, single-user. No login. The page, its SSE stream (/live), and the plan-review reply are reachable over loopback/tailnet (_PUBLIC_PATHS); the dangerous surface (WebRTC offer, /harness, /start, /client) stays behind the RELAY_VOICECALL_BEARER gate with a loopback exemption. This is byte-identical to the pre-auth cockpit — zero config.
• hosted — multi-tenant. A login is required: the tailnet carve-out is gone, every route except /healthz + /login needs a session cookie, and each account is scoped to its own ~/.yothere-<tenant> home (separate threads/data/state), so two logins see two isolated fleets with no cross-tenant read. Auth is Python-native and stdlib-only (yothere.voicecall.auth): scrypt password hashing, an opaque session token in SQLite (RELAY_AUTH_DB, WAL, separate from fleet state), login lockout, and a Secure cookie tied to the mode (not the request URL, so a TLS-terminating proxy can't drop it). Tenant homes live under RELAY_TENANTS_ROOT (default: beside ~/.relay).

Hosted env: RELAY_AUTH_MODE=hosted, RELAY_AUTH_DB=<path>, RELAY_TENANTS_ROOT=<dir>. This mode is deployed at app.yothere.ai (Fly + managed Postgres) with invite-gated signup. Execution is BYO-compute: each tenant pairs their own machine, which long-polls a device-token job queue and runs the brain locally, so advancing a tenant's threads no longer needs a server-side per-tenant runner. A control-plane supervisor tick handles cross-tenant housekeeping (lease expiry, ask-park). Open (non-invite) public signup and a psycopg connection pool are the next increments.

MCP surface

Drive a yothere fleet from any MCP client (Claude Desktop, etc.) — an orthogonal channel that exposes the same front doors as the CLI and cockpit over stdio MCP:

• spawn_thread(task, mode, focus) — spawn fleet thread(s) from a natural-language task (read/draft-only, blocks for your approval before any outward action).
• board() — the fleet at a glance, needs-eyes first.
• reply(thread_id, text) — approve / edit / reject a blocked thread.

Single-user / local (drives this machine's ~/.relay fleet). mcp is an optional extra; install + register:

pip install 'yothere[mcp]'

// 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

v1.5.6 — invite-only beta. The fleet runner, board, voice surface, and the remote-brain path are tested (python tests/relay_test.py). The hosted multi-tenant cockpit is live at app.yothere.ai: invite-gated signup, BYO-compute (your laptop runs the brain via a device-token job queue while a hosted control plane holds auth + state), and hosted voice (Gemini-Live over a laptop-hybrid broker with cross-NAT TURN, in beta). See CHANGELOG.md for the increment log.

Contributing

Two people build yothere and neither should break main alone. Every change goes through a PR that the other founder approves before it merges — see CONTRIBUTING.md.

License

MIT — see LICENSE.