Hallow logoHallow Agent

Local-first autonomous runtime

Hallow Agent Documentation.

Hallow is an installable agent OS for builders who want autonomous agents to run locally, remember work privately, discover tools through standards, connect to models, and improve skills under policy. This is the docs surface: direct project documentation, not community claims.

Diagram of the Hallow local runtime connected to memory, tools, models, gateways, skills, and audits.
Local runtime map: memory, tools, models, gateways, signed skills, and audit checks around one Hallow process.
Local runtime

One process serves the desktop shell, API, scheduler, memory, tool registry, and readiness checks.

Private memory

Agents store traces, decisions, vectors, corrections, and project context inside the local Hallow home.

Standard agents

Agents and skills ship with manifests, permissions, model needs, signatures, and audit metadata.

Quickstart

The public flow stays small: one Windows command that works from CMD or PowerShell, then hallow. Everything advanced lives behind the terminal.

Four step Hallow boot sequence: install, setup, doctor, start.
The first-run path stays simple before deeper runtime controls appear.
powershell -nop -ep bypass -c "irm https://hallow-agent.xyz/install.ps1|iex"
hallow
hallow> status
hallow> start
hallow> run "summarize this workspace"
Use the command above directly in CMD or PowerShell. macOS/Linux commands are listed below for shell-specific installs.

Installation

01
Windows CMD / PowerShell
powershell -nop -ep bypass -c "irm https://hallow-agent.xyz/install.ps1|iex"
02
macOS / Linux / WSL2
curl -fsSL https://hallow-agent.xyz/install.sh | bash
03
PowerShell-native fallback
irm https://hallow-agent.xyz/install.ps1 | iex
04
Android / Termux

Use the Linux installer inside Termux after Node and Git are available. Mobile support is a runtime target, not the recommended first install path.

Updating & Uninstalling

Update

Pull the newest project version, rebuild packages, then run hallow setup to regenerate desktop artifacts.

Backup

Keep a copy of the Hallow home folder before destructive changes. It contains memory, traces, config, packages, and audits.

Uninstall

Stop the runtime process, remove the installed command, then archive or delete the Hallow home folder.

Learning Path

1. Run the runtime

Open /desktop, check readiness, and understand the local boundary.

2. Create an agent

Generate a manifest, define capabilities, attach model profiles, and keep permissions explicit.

3. Add tools

Connect MCP, browser, OAuth, and gateway adapters only after policy checks pass.

4. Package skills

Turn repeated workflows into signed skills with tests and metadata.

5. Enable autonomy

Use scheduler, quality scoring, traces, and repair loops before unattended runs.

6. Publish safely

Export package metadata, audit security, and avoid shipping private local paths.

Using Hallow

Hallow has three visible surfaces: CLI for builders, desktop shell for local operation, and HTTP APIs for automation.

Local Hallow memory vault storing traces, decisions, and vector items.
Memory stays inspectable inside the local Hallow home: traces, vectors, notes, and repair context.
Agents

Autonomous workers described by manifests, memory policy, model profiles, tools, safety limits, and trace export rules.

Memory

Local vault for semantic memory, run traces, corrections, and future repair context.

Models

OpenAI-compatible APIs, Claude/OpenRouter-style providers, local fallback, and cost-aware route profiles.

Gateway lanes

CLI, browser shell, chat adapters, and device lanes with pairing and allowlist controls.

Features

Signed Hallow package metadata with permissions, model needs, checksum, and audit status.
Agents and skills are not loose prompt folders; they move through package metadata, signatures, and audit checks.
MCP tools

Register stdio or HTTP MCP servers and filter tools per server before an agent can call them.

Skills

Reusable workflows with metadata, tests, permissions, checksums, and optional external skill directories.

Marketplace

Signed packages for agents and skills, with trust metadata and install verification.

Self-healing

Failed runs become traces, traces become repair candidates, and quality gates decide promotion.

Security audit

Checks exposed secrets, unsafe permissions, sandbox posture, package trust, and gateway policy.

Desktop shell

A clean local UI for install flow, readiness checks, and runtime status without leaking local paths.

Integrations

Hallow gateway lanes connecting local agent to CLI, browser, Telegram, Discord, Slack, and devices.
Gateway lanes let one local agent reach multiple channels while policy stays in front.
Model providers

OpenAI-compatible APIs, OpenRouter, Anthropic-style routes, and local model profiles.

OAuth

Connector definitions and grants are separated from UI display so hidden tokens are not exposed.

Browser

Live browser sessions can capture HTML, screenshots, and artifacts for traceable tasks.

Messaging platforms

Gateway adapters are the path for Telegram, Discord, Slack, WhatsApp-style channels, and custom devices.

Guides & Tutorials

Create first agent

hallow agent create research
hallow doctor
hallow start

Package a skill

hallow package sign ./skills/research-briefing
hallow marketplace verify --path ./skills/research-briefing

Connect MCP server

hallow mcp add filesystem --command npx --args -y,@modelcontextprotocol/server-filesystem,.
hallow mcp probe --server filesystem

Pair a gateway

hallow gateway pair telegram --allowlist team.yaml
hallow gateway status

Developer Guide

Manifest standard

Agents and skills declare identity, capabilities, permissions, model needs, memory policy, and trace export rules.

Runtime API

Local HTTP endpoints expose health, readiness, desktop status, tasks, memory, tools, marketplace, OAuth, and gateways.

Adapter contract

Gateway and tool adapters must declare transport, allowed actions, auth policy, and failure behavior.

Tracing

Every meaningful autonomous action should leave an artifact: task record, browser capture, tool event, memory item, or audit row.

Reference

Core commands

hallow terminal, hallow setup, hallow doctor, hallow start, hallow agent create, hallow mcp add.

Files & folders

Hallow home stores config, memory, traces, packages, marketplace indexes, desktop artifacts, audits, and browser sessions.

Troubleshooting

If a route returns 404, rerun hallow setup and restart the runtime so generated desktop/docs artifacts are refreshed.