The builder · concept preview

Build your agent.
Part by part.

Every ability is a sealed capsule: the brain, the memory, the tools, the guards. Snap them together, then swap any part later without rebuilding the rest. Your agent can never take more than you gave it.

Concept preview. The builder is not live yet: assembling and installing an agent from this page is coming, and a build made here would break until the backend lands. What is real today, and shown honestly below: every capsule is on GitHub now, the manifest is a real Distro.toml, and astrid init --distro installs one. Until the builder ships, install Unicity Astrid OS and run astrid init.

Step 1 · name it, and pick where it lives

agent name

wanderer

Standalone daemon Claude Code Codex soon Agent Hermes soon OpenClaw soon

Standalone runs as its own daemon: every frontend is an uplink to the same agent. On Claude Code, Claude is the brain and the frontend, so those two slots drop away.

Step 2 · start from an archetype

⌨ The Coder Ships code. Files, shell, the forge, and it remembers the project.

🔍 The Researcher Reads everything, remembers everything, cites its sources.

🎩 The Concierge Runs on your GPU, reachable from Telegram.

⚙ Bare Metal Just the loop and a way in. You add every capability yourself.

Step 3 · the workbench

◆

The brain

required

OpenAI-compatible LM Studio, Ollama, llama.cpp, on your GPU OpenAI GPT models via your OpenAI key

↻

The loop

required

ReAct The stateless reason-and-act coordinator

▤

The memory

optional

Cross-session memory Reads .astrid/memory.md into the prompt key: files Session store Persistent, append-only conversation history

⚒

The toolbelt

pick any

Files Read and write your project, behind the VFS airlock key: files Web Fetch and call APIs, behind the SSRF airlock key: web Shell Run commands, gated behind approval key: commands Self-manage Inspect and manage its own runtime

✦

The skills

optional

Skills loader Loads SKILL.md packs from the workspace key: files Project rules Injects AGENTS.md into the prompt key: files The forge Teaches the agent to build its own capsules

⌗

The frontends

need one

Terminal The TUI, bridged over a Unix socket Telegram A Telegram bot as a frontend

your agent

◆ OpenAI-compatible brain

↻ ReAct loop

▤ Cross-session · Session memory

⚒ Files · Web · Shell · Self-manage tools

✦ Skills · Rules · Forge skills

⌗ Terminal frontend

the engine · dumb on purpose

wanderer

8 capsules · files + web + commands

keys it will hold

✓ files✓ web✓ commands passwords

Keys only ever get fewer down the chain. Passwords stay in the secret store and are never handed to the model.

Distro.toml real format

# assembled at astridos.org — a starting point, edit freely
schema-version = 1

[distro]
name = "wanderer"
description = "Assembled with the Astrid builder"

[[capsule]]
name = "openai-compat"
source = "unicity/openai-compat"
group = "llm"
role = "provider"

[[capsule]]
name = "react"
source = "unicity/react"
role = "coordinator"

[[capsule]]
name = "memory"
source = "unicity/memory"

[[capsule]]
name = "session"
source = "unicity/session"

[[capsule]]
name = "fs"
source = "unicity/fs"

[[capsule]]
name = "http"
source = "unicity/http"

[[capsule]]
name = "shell"
source = "unicity/shell"

[[capsule]]
name = "cli"
source = "unicity/cli"
role = "uplink"

Step 4 · take it home

Run it as its own daemon

The builder would hand you these. Save the Distro.toml, then point astrid init at it: it installs every capsule, prompts for any keys, and writes a per-agent lockfile.

brew install unicity-astrid/tap/astrid macOS and Linux, one daemon
astrid init --distro ./wanderer.toml assembles wanderer from the manifest above
astrid distro apply -a wanderer the distro lockfile is per-principal

Want to build the parts instead of the agent? The playground runs the real kernel and lets you write a capsule in this tab, today.