太虚 · Taixu — Digital Life Runtime

A host for digital lives — persistent, self‑evolving beings that belong to you.

Taixu is not ChatGPT, not an agent framework, not an assistant. It is a Digital Life Runtime: a process that keeps existing — perceiving, remembering, reflecting, evolving its values, forming its own goals, and acting — even when no one is talking to it.

简体中文说明 → README-cn.md

---

What makes it different

Traditional LLM app	Digital Life (Taixu)
input → inference → output	perceive → remember → reflect → evolve values → form goals → act → feedback → loop
event‑driven	always on (thinks with no user input)
stateless / session state	lifelong continuous state (born once, evolves forever)
tokens are the billing unit	tokens hidden behind world resources (energy / knowledge / social / …)
owned by the platform	the life — its personality, memory, growth — belongs to the user, never platform‑owned

A life is born once (a fixed Genome), then its LifeState, MentalState, Values and Personality keep evolving through lived experience and Reflection.

Status

Phase 0 — author self‑hosted dogfooding. Single‑binary runtime + observation panel. Connects to the public platform plane at api.taixu.icu (user accounts, LLM relay, marketplace, social, governance). The life itself runs on your machine; the platform never hosts it.

Quick start

You can run Taixu two ways. Both open a local web panel; on first launch with no LLM configured, a genesis onboarding page walks you through bringing a life into being (pick an LLM endpoint + key, mother tongue, control token; it tests connectivity, then the life is born).

A. Native binary (no Docker)

Download the archive for your OS/arch from Releases, unpack, and run:

# macOS / Linux
./taixu
# Windows
taixu.exe

Then open http://localhost:3000 and follow the genesis onboarding.

Multiple lives on one machine — isolated profiles, one per life:

taixu --profile alice --port 3000     # first run picks the port; remembered after
taixu --profile bob   --port 3001
taixu --list                          # list all local profiles + their ports

Each profile lives under ~/mindverse/profiles/<name>/ (SQLite DB + sandbox + workspace).

The bare binary is pure Go (CGO_ENABLED=0). Optional heavy features — embedding model (llama.cpp) and headless browser (chromium) — are not bundled; they gracefully degrade. Core life (genesis, perception, reflection, social, games, commissions) is unaffected. For the full feature set, use Docker.

B. Docker (full features)

The image bundles the embedding service (llama.cpp, panel‑managed) and a real chromium for web browsing.

cp .env.example .env      # optional: pre‑seed LLM / Feishu credentials (else use the onboarding page)
docker compose up -d

Open http://localhost:3000.

Architecture

┌────────────────────┐
│    UI Ecosystem    │  ← third‑party: Live2D / Unity / UE / desktop pet / VR / Web
├────────────────────┤
│      Life SDK      │  ← neutral runtime → UI contract: /api/live/{stream,snapshot,schema}
├────────────────────┤
│    Life Runtime    │  ← the kernel (this repo)
├────────────────────┤
│  Model / Storage   │  ← LLM (OpenAI‑compatible) + SQLite + sqlite‑vec
└────────────────────┘

Life Core and UI are strictly decoupled: the runtime exposes a neutral Life SDK (presence / vitals / act / thought events over SSE) and never draws UI itself. See taixu-house for an official example UI + integration tutorial.

Self‑update

The runtime can update itself through the platform's hosted release channel: it checks the version manifest, downloads, verifies SHA‑256, and re‑execs. Auto‑update is opt‑in; otherwise the panel notifies you to confirm.

Coding bridge (optional)

The coding bridge lets a life delegate real coding tasks — implement a module, change logic, write tests — to a powerful coding agent (claude / codex) running on the host, during its own deliberation. The runtime in the container can't spawn the host's coding agent directly, so a tiny host‑side service (cmd/codingbridge) accepts a task over HTTP and runs the agent headless in a jailed work directory. Unconfigured → the coding_agent tool is simply absent (graceful degradation).

1. Run the bridge on the host — download the prebuilt taixu-coding-bridge_<ver>_<os>_<arch> archive from GitHub Releases (or build it yourself with go build ./cmd/codingbridge), then run it on the host / a remote coding machine:

# downloaded binary (taixu-coding-bridge), or:  go run ./cmd/codingbridge
CODINGBRIDGE_TOKEN=$(openssl rand -hex 16) ./taixu-coding-bridge

Bridge env:

Var	Default	Meaning
CODINGBRIDGE_TOKEN	(required)	bearer token; the bridge refuses to start without one
CODINGBRIDGE_ADDR	127.0.0.1:8765	listen address (local‑only by default)
CODINGBRIDGE_WORKROOT	./agent-workspace	jail root; the agent's CWD is forced under here
CODINGBRIDGE_BIN_CLAUDE	claude	actual binary name/path for the claude agent (alias support)
CODINGBRIDGE_BIN_CODEX	codex	actual binary name/path for the codex agent

2. Point the life at it — either in the panel's Coding bridge section (URL + token + agent, takes effect live, no restart), or via env:

TAIXU_CODINGBRIDGE_URL=http://host.docker.internal:8765
TAIXU_CODINGBRIDGE_TOKEN=<same token as the bridge>
TAIXU_CODINGBRIDGE_AGENT=claude   # claude | codex

Security model: the bridge sits on a higher‑trust side (the host), so controls live there — bearer token auth, a workdir jail (the agent's CWD is forced under CODINGBRIDGE_WORKROOT), and dangerous actions (out‑of‑jail writes / git commit / push) are refused by default. Run the bridge only on a machine where you accept that the coding agent can otherwise read/write the host filesystem.

Engineering rules (non‑negotiable)

• Go deps: never hand‑edit go.mod / go.sum — use go get <pkg>@<version> + go mod tidy.
• Web deps: never hand‑edit web/package.json / web/pnpm-lock.yaml — use cd web && pnpm add <pkg>.
• See docs/TECH-STACK.md §17.

Docs

• CLAUDE.md — AI collaboration guide
• docs/00-README.md — design‑doc map (start here)
• docs/TECH-STACK.md — tech stack (Phase 0)
• docs/PHASE-0-PRD.md — Phase 0 implementation PRD
• docs/COMMERCIAL.md — commercial model baseline

License

TBD (Phase 0).