Agent Hub (ARW)

Your private AI control room that can scale and share when you choose.

In plain terms: Agent Hub (ARW) lets you run your own team of AI “helpers” on your computer to research, plan, write, and build—while laying the groundwork for upcoming voice and vision helpers—all under your control. It is local‑first and privacy‑first by default, with the option to securely pool computing power with trusted peers when a project needs more muscle.

ARW now anchors four pillars: (1) cultivate empathetic personas (preview) with transparent self-models and vibe feedback, (2) deliver infinite-context memory overlays on any local or remote model, (3) unlock a fair autonomous economy so anyone can earn with policy-aligned agents on consumer hardware, and (4) guarantee universal access through eco defaults, offline bundles, and sustainability safeguards so every laptop can participate for free.

Restructure update: arw-server is now the sole API surface (headless-first) across every deployment. The old bridge layer and its launch flags have been retired in favour of the unified stack.

Full documentation → Docs home Assistant quickstart → Agent Onboarding

Quick Links

• Agent Onboarding - Workflow primer for assistants and new contributors.
• Quickstart - Run the unified server locally.
• Quick Smoke - Fast guardrail (verify --fast) to confirm the toolchain before full setup.
• Runtime Quickstart - Non-technical checklist for preparing managed runtimes (zero-auth mirrors + checksum validation).
• Persona Preview Quickstart - Enable the optional persona stack and seed your first preview persona safely.
• Universal Access Starter Kit - Generate an offline-ready bundle (eco presets, docs, starter persona) for low-spec installs.
• Orchestrator CLI - Catalog, launch, and monitor persona-aware training runs from the terminal.
• Repo Map - Directory overview for retrieval and navigation.
• Documentation Home - Product overview, guides, and reference docs.
• CLI: Tail events quickly — arw-cli events tail --prefix service.,state.read.model.patch --replay 25 --store .arw/last-event-id (see docs/reference/cli.md)
• Grafana: SSE row snippet — import docs/snippets/grafana_sse_row.json into a dashboard
• Adapters quick smoke (one-shot): just adapters-smoke-oneshot or mise run adapters:smoke:oneshot

Adapters

• Gallery: see sample manifests and quick commands at docs/guide/adapter_gallery.md.
• SDK: build custom adapters with crates/arw-runtime-adapter — guide at docs/guide/adapter_sdk.md.
• Validate: docs/guide/adapters_validate.md (single-file, lint-all, changed-only).

Release Status

• Legacy release bundles (<= v0.1.4) have been retired; no prebuilt downloads are published while we refocus on the next milestone.
• The main branch now tracks the 0.2.0-dev pre-release; GitHub Releases remain frozen until the next milestone cut.
• Prefer source builds (or local packaging scripts) for testing and development during the pre-release cycle.

Getting Started

• Clone the repo and install Rust 1.90+ via rustup.
• Validate your toolchain quickly with the Quick Smoke flow (bash scripts/dev.sh verify --fast) before installing optional stacks.
• Prebuilt installers are currently unavailable; build locally with scripts/dev.{sh,ps1} setup or produce a bundle via scripts/package.{sh,ps1} if you need a portable archive.
• For source builds, run scripts/dev.{sh,ps1} setup (defaults to headless + non-interactive -Yes) or follow docs/guide/quickstart.md; pass --with-launcher / -WithLauncher when desktop UI prerequisites are ready.
• Prefer an all-in-one toolchain bootstrap? Install mise and run mise install to pin Rust 1.90, Python 3.12, Node.js 18, jq, and ripgrep; mise run verify, mise run verify:fast, and mise run verify:ci wrap the guardrail helpers.

Build & Test

• Native Windows flow: run scripts\dev.ps1 from PowerShell; switch to WSL only when you need Linux-only tooling.
• Unified helper: scripts/dev.{sh,ps1} wraps the common workflow (e.g., scripts/dev.ps1 build, scripts/dev.ps1 verify).
• Guardrail sweep: scripts/dev.sh verify (headless default skips the Tauri crate; pass --with-launcher or set ARW_VERIFY_INCLUDE_LAUNCHER=1 to include it). Just validating your Rust toolchain? Start with Quick Smoke for scripts/dev.sh verify --fast. Use --ci for a CI-parity sweep (registry integrity, docgens in --check mode, env-guard, and smokes). Missing Node.js auto-skips the launcher UI smoke; set ARW_VERIFY_REQUIRE_DOCS=1 to require Python/PyYAML for doc checks.
• TypeScript client: when Node.js and npm are available and you do not pass --fast, verify also builds the minimal TypeScript client and arw-events bin under clients/typescript (skips automatically if Node/npm are missing).
• Build: scripts/build.sh (Linux/macOS) or scripts/build.ps1 (Windows). Both default to a headless build that skips the Tauri launcher; add --with-launcher / -WithLauncher or set ARW_BUILD_LAUNCHER=1 when you need the desktop UI (requires WebKitGTK 4.1 + libsoup3 on Linux or WebView2 on Windows). make build / just build follow the same headless default, with make build-launcher / just build-launcher opting into the full workspace build.
• Format: cargo fmt --all
• Lint: cargo clippy --workspace --all-targets -- -D warnings
• Tests: cargo nextest run or scripts/test.{sh,ps1}
• Runtime smoke: just smoke-safe (apply low-impact defaults) then just runtime-smoke. The suite always runs the stub stage, flips CPU/GPU policies to auto automatically once you set RUNTIME_SMOKE_ALLOW_CPU=1 or RUNTIME_SMOKE_ALLOW_GPU=1, and will auto-download TinyLlama weights when HF_TOKEN/HUGGINGFACEHUB_API_TOKEN is exported (disable with RUNTIME_SMOKE_SKIP_AUTO_WEIGHTS=1). Provide LLAMA_SERVER_BIN/LLAMA_MODEL_PATH for real coverage; otherwise the run falls back to simulated markers.
• Docs: mkdocs build --strict, scripts/dev.{sh,ps1} docs, or just docs-build (requires Bash). Windows without Bash: pwsh -ExecutionPolicy Bypass -File scripts\docgen.ps1 followed by mkdocs build --strict. Prefer task wrappers? Use mise run docs:check or mise run docs:check:fast.
• Consent validation: run python3 scripts/validate_runtime_consent.py whenever you edit configs/runtime/bundles*.json so audio/vision bundles keep their metadata.consent annotations (CI blocks merges otherwise).
• Docs metadata: python scripts/update_doc_metadata.py docs/<path>.md refreshes the Updated: stamp after edits (add --dry-run to preview changes).
• Docs toolchain: mise run bootstrap:docs (or bash scripts/bootstrap_docs.sh) installs the pinned MkDocs/Material stack. Generate an offline wheel bundle with mise run docs:cache:build (archive lands in dist/docs-wheels.tar.gz), then reuse it via mise run bootstrap:docs -- --wheel-dir <path>.
• Need offline docs? Generate docs-wheels.tar.gz locally with mise run docs:cache:build (or scripts/dev.sh docs-cache / scripts\dev.ps1 docs-cache) and reuse it via scripts/bootstrap_docs.sh --wheel-dir <dir>.

Repository Layout

• crates/ - Core Rust libraries (protocol, kernel, policy, runtime, etc.).
• apps/ - arw-server, CLI, connectors, and launcher surfaces.
• docs/ - MkDocs site content; published via GitHub Pages.
• spec/ - JSON Schemas and machine-readable contracts.
• scripts/ - Cross-platform helpers for setup, testing, and packaging.
• See docs/ai/REPO_MAP.md for more detail and additional directories.

Architecture & Roadmap

• Unified stack details live in docs/RESTRUCTURE.md (source of truth for the migration).
• Feature status and planned work are tracked in docs/ROADMAP.md and docs/reference/feature_matrix.md.
• Release notes land in CHANGELOG.md and RELEASE_NOTES_* files.

Contributing

• Review CONTRIBUTING.md, CODE_OF_CONDUCT.md, and the Agent Onboarding guide before opening changes.
• Follow the assisted development workflow (PLAN -> DIFF -> tests) outlined in docs/ai/ASSISTED_DEV_GUIDE.md.

Licensing

• Dual licensed under MIT and Apache-2.0.
• Security disclosures: see SECURITY.md.

Community & Support

• Issues and discussions live on GitHub.
• Automation dashboards and task tracking sit under .arw/ (large generated files are excluded from releases).
• For status badges, CI pipelines, and deployment artifacts, see the badges at the top of this README.