Overview Capabilities Product Memory Smart home Security Get started Privacy
中文 View on GitHub
0.2.0 preview

A home agent that lives on your computer

EnvoyHome is a self-hosted, security-first home agent: chat, memory, a document knowledge base, static workflows, approvals, and optional smart-home control — all running on your own machine. Your desktop and phone are just windows onto it.

Agent loop Workflows Memory L0–L7 Doc KB Approvals Home Assistant MQTT Skills MCP Schedules
8
Memory layers · L0–L7
6
First-party chat channels
2
Thin clients · desktop + phone
MIT
Open source · self-hosted
EnvoyHome architecture — desktop and phone thin clients dial the home agent daemon over home.*; the core is envoy-harness, workflows, memory, approvals, and the tool catalogue; Home Assistant and MQTT hubs are optional

Thin clients dial the home agent over home.*. The center is the agent loop, workflows, memory, and tools; hubs are optional.

What is EnvoyHome?

A household agent you own — not a cloud chatbot with a home-themed name.

EnvoyHome is the household agent in the EnvoyMesh apps family. The core is the agent loop itself: envoy-harness, static workflows, memory, approvals, and a rich tool catalogue the agent can use — files, web, browser, office documents, skills, schedules, MCP, and more.

The agent lives on the home computer. Keys, transcripts, sandboxes, and smart-home credentials stay there. Desktop and phone dial in over the home.* channel; they are windows, not brains.

Smart home is one tool surface among those capabilities — useful when you connect a hub, never the definition of the product. EnvoyHome enhances Home Assistant and MQTT with natural language, approvals, and workflows. It does not replace your hub, and it does not ship a social network.

🏠

The agent stays home

Every chat, workflow, and scheduled turn runs through the home daemon. Thin clients reconnect to it — they never hold the model keys or run the loop.

🔁

Workflows are not a second engine

Static DAGs in a visual diagram or YAML, drafted from natural language. Workflow nodes call the same tools as chat, with the same approvals.

🛡️

Approvals before consequences

Durable pending approvals, a 600-second default TTL, fail-closed decisions, and grants bound to the resolved call — not a blanket “agent may do anything” switch.

🧠

Memory that learns carefully

Profile, standing notes, daily logs, episodic search, KnowYou, a document knowledge base, and skills — with PendingLearn gating the durable writes.

👨‍👩‍👧

One agent, many people

Household profiles share the agent but keep isolated memory, files, and policy. Phones pair to the house, and each device is bound to the accounts it may act for.

🔀

Cloud, local, or careful mix

Bring an OpenAI-compatible or Anthropic-compatible provider, run llama.cpp locally, or connect Ollama. Privacy Mode forces local turns and can block smart-home egress.

🧰

A real tool catalogue

Files behind a path jail, web search and reading, a browser with form cards, office document import/export, skills, sessions, MCP servers, and optional HA/MQTT tools.

🧩

Skills pass a verify gate

The on-disk format is a SKILL.md directory. Install from disk, a URL, or ClawHub — then verify capabilities before the skill can run. Agent-authored skills wait in PendingLearn.

⏰

Schedules, heartbeat, standing orders

Repeating or one-shot jobs, natural-language “when” proposals you confirm, a heartbeat for periodic review, and standing orders that survive restarts.

A Settings-first product

The desktop shell is where the household is configured, the agent is watched, and approvals are answered.

EnvoyHome desktop — Connections → Home Assistant with demo hub, device binding, and safety options

Connections → Home Assistant

Connect a hub or start the bundled demo. Devices must be bound before the agent can use them; safety locks always ask first.

EnvoyHome desktop — Chats with the session rail, profile switcher, and inline approval surface

Chats

A session rail per profile, a composer for the home agent, and approvals shown inline — Allow, Deny, or narrow to this exact call.

Captured from the local desktop Settings shell against a demonstration profile. Click any screenshot to enlarge.

Your home agent, from your phone

The EnvoyHome Flutter preview for iOS and Android is a thin client for a daemon you run. Pair once, then chat, answer approvals, and manage the same home setup over your own connection.

01

Pair without a cloud account

Scan the desktop QR, paste a pairing link, enter host:port, or connect over SSH. The phone dials your home daemon directly.

02

Chat, approve, and share

Profile-aware conversations, inline Allow / Deny approvals, attachments, and OS dictation. Shared files and artifacts use the same home connection.

03

One household control plane

Chat, a Schedules hub (Jobs · Scenes · Histories), and Settings mirror the desktop. LLM, Connections, Skills (packs), and System stay in one Settings hub.

EnvoyHome mobile app — Connect screen with empty pairing link field and QR, host:port, and SSH options

Pair your way

The real Connect screen with Link, host:port, and SSH options. The pairing field is intentionally empty.

EnvoyHome mobile app — a generic morning routine conversation with a real assistant reply

Chat on the go

A real, generic exchange over the paired daemon connection. The conversation contains no personal or household data.

EnvoyHome mobile app — connected Settings hub with LLM, Connections, Skills, and System

One connected Settings hub

LLM, Connections, Skills, and System mirror the same desktop control plane over the paired home connection.

Preview thin client (0.2.0). A reachable EnvoyHome daemon is required; the phone app is not a standalone cloud chatbot. The optional Home tab appears only when Home Assistant is configured and enabled. APNs / FCM push is optional (approvals and schedule notifies) and configured on the daemon. Store listing pack is prepared; there is no App Store or Play listing live yet.

Captured from the actual Flutter app on an iOS simulator. Click any phone screenshot to enlarge.

Eight layers, one honest memory

EnvoyHome does not treat “everything is memory” as a design. Each layer has a job, a budget, and an injection rule — so the agent can remember without quietly rewriting who you are.

  • Working memory is the session transcript, flushed before compaction rather than silently truncated.
  • Standing memory is small and capped: profile facts, MEMORY.md, and the last two days of notes.
  • Lookup memory is a tool: episodic search, KnowYou, and your document knowledge base with citations.
  • Durable writes go through PendingLearn — you accept or reject the proposed diff before L1, L2, or a skill changes.
L0
Workingsession transcript
flushed
L1
Profilepreferences, identity
capped
L2
Standing notesMEMORY.md
capped
L3
Dailydated observations
today + yesterday
L4
Episodic searchsession FTS
tool
L4′
KnowYouderived episodic facts
budgeted
L6
Doc KBknowledge/ files + index
budgeted or tool
L7
Skillsprocedural “how”
progressive

L5 (deep store) is a slot: files by default, heavy RAG backends only as optional plugins. Files remain the source of truth.

The document knowledge base is yours. Drop files under knowledge/; EnvoyHome indexes them per account in local SQLite and answers with citations. Hybrid search can use the Home embed engine on :18793 — all on the home machine. Your documents are not uploaded to an EnvoyHome cloud, because there isn’t one.

Smart home, with a safety class

Connect Home Assistant or an MQTT broker and the agent can read state and — with approval — call services. It integrates with your hub; it does not become one.

  • Actuation is admin, not a file write. Unattended turns need a matching grant; script, scene, and button indirection is denied by default.
  • Safety-class calls never ride a grant. Locks, alarms, garage doors, gates, valves, unlisted MQTT topics, and anything marked neverUnattended ask every single time.
  • Unknown means safety-class. A new domain, a vendor extension, or a topic you did not allow-list fails closed instead of guessing.
  • Bind before control. Objects are registered by the authenticated hub and resolved through a chokepoint — never from a plugin-asserted id.

Preview posture: Home Assistant REST tools and MQTT publish work today, with bundled demo hubs for the full Chat → Approvals path. Live inbound event plumbing and a signed-off live household still follow the operator checklist.

EnvoyHome Connections — Home Assistant tab with demo hub, device type picker, always-ask option, and per-home bindings

Reach it from where you already are

The EnvoyMesh thin client is the primary path. First-party chat plugins cover six common platforms, and an HTTP hatch takes care of the rest.

EnvoyMesh thin clients Primary

Desktop and phone dial the house directly over home.*, with full Settings and Chat. Pair once by QR, host:port, or SSH hop.

Telegram Chat

The reference bot plugin. Bind a sender to a household profile; replies stay on the same thread.

Discord Chat

DM-based chat through the official bot API. No guild-wide actuation.

Slack Chat

Socket Mode app with bot and app tokens. Direct messages only.

Feishu / Lark Chat

Long-connection events, p2p messages only — no public callback URL required.

DingTalk Chat

Stream Mode robot for 1:1 conversations.

WeCom Chat

WeChat Customer Service sync for 1:1 support threads.

HTTP inbound hatch Escape hatch

A minimal /v1/inbound endpoint for bots that never load as plugins.

Chat plugins ship with setup guides and a shared sender → profile link. Live bot verification is still in progress for the IM set — for preview use, the phone thin client is the primary path.

Security is the product, not a page

Shared-home defaults: sandbox on, approvals durable, unknown things fail closed.

Identity

No account, no cloud mailbox

Identity is an Ed25519 key pair on your devices. There is no EnvoyHome sign-up, no email address, and no hosted transcript store to breach.

Approvals

Durable and fail-closed

Pending approvals and grants are files per account. A restart keeps the question; a timeout resolves to deny; an unresolvable principal is denied.

Execution

Path jail and risk tiers

File tools resolve through one jail. Tools declare read, write, exec, network, admin, or sensitive — and can escalate, never de-escalate.

Privacy

Privacy Mode

A hard switch that forces local turns and refuses cloud tools. With smart-home egress blocking on, it also refuses actuation and event-source notifications.

Accounts

Household isolation

Memory, transcripts, files, grants, and policy are keyed by account. A paired device may only act for the accounts it is bound to.

Smart home

The physical world is different

Actuation is admin. Camera, lock, and presence data is sensitive and forces local inference. Device state never becomes standing memory automatically.

Read the privacy policy →

EnvoyMesh Family

Three apps sharing one mesh — identity, pairing, and transport — while each keeps its own brain and its own state.

EnvoyMesh logo

EnvoyMesh

The desktop home node and peer-to-peer mesh that the rest of the family is built on.

Visit envoymesh.cn →
EnvoyDev logo

EnvoyDev

The control plane for coding agents — reach your own machines from a window or a phone.

Visit envoymesh.cn/envoydev →
Veda Notes logo

Veda Notes

A local-first AI journal that pairs with the home node for private sync and home AI.

Visit envoymesh.cn/veda-ai →

Run it at home

Preview builds run from source. You need Node.js 22+, pnpm, and a sibling EnvoyMesh checkout.

1

Check the siblings

EnvoyHome links @envoymesh/* with link: — they are not on npm. The peer check fails closed if a sibling is missing or on the wrong branch.

./scripts/peers-check.sh
# or set ENVOYHOME_PEERS_GIT_BASE and run ./scripts/fetch-peers.sh
2

Install and build

Never install with CI=true set — pnpm’s strict-mode trap is a documented convention in this repo.

env -u CI pnpm install
pnpm -r build
3

Start the home agent

Run the daemon alone, or let the desktop shell supervise it. Then create a profile, add a model, and pair a phone when you want one.

pnpm run dev:daemon   # daemon only
pnpm desktop:dev      # Tauri Settings shell
WebSocketws://127.0.0.1:4780/ws
Healthhttp://127.0.0.1:4781/health

Source and issues

Scan to open the repository, or use the link below.

GitHub repository

Platforms

Desktop shellmacOS, Linux, Windows
PhoneiOS + Android
Local modelsllama.cpp · Ollama
Home embed:18793
LicenseMIT

macOS is the primary desktop target; Linux and Windows are best-effort. The phone app is a Flutter preview.

EnvoyHome 0.2.0 is an early-adopter preview. v1.0.0 is not tagged yet, and there is no App Store or Play listing live — install from source or a locally built package under release/.

Run the desktop app from source

Node.js ≥ 22 and pnpm 10.33 (the root packageManager), a sibling EnvoyMesh checkout, and Rust plus the Tauri CLI prerequisites for the desktop shell.

./scripts/peers-check.sh
env -u CI pnpm install
pnpm -r build
pnpm desktop:dev        # daemon build if missing, then Tauri dev

desktop:dev runs ./scripts/desktop-dev.sh: it builds packages/daemon/dist/bin/envoyhome-daemon.js only when that file is missing, then starts tauri:dev. Tauri supervises the daemon once the binary exists. No provider config file is needed before launch — first-run Settings asks for a profile and a model provider (preset + key, or Local / Ollama).

Release installers: ./scripts/build-desktop.sh macos → release/envoyhome-desktop-0.2.0-macos-*.dmg (Developer ID + notarization required; push credentials staged unless REQUIRE_PUSH_CREDENTIALS=0). Windows: .\scripts\build-desktop.ps1 on a Windows host.