hybridclaw

HybridClaw

Business Skills That Work On Smaller Models

HybridClaw's main promise is practical business work. The bundled CRM, marketing, analytics, finance, cloud, office, and operations skills are implemented as tested helpers with eval scenarios, credential boundaries, and approval tiers, and are validated against Qwen/Qwen3.6-27B-FP8 as the small-model baseline.

That is the useful difference: a compact model can operate Salesforce, HubSpot, GA4, Google Ads, Lexware, Airtable, Miro, Zoho, Microsoft 365, warehouse SQL, invoices, cloud ops, and office documents through structured rails instead of fragile free-form prompting.

HybridClaw also treats agents as networked coworkers. Local agents, hosted HybridAI proxy agents, and trusted peer HybridClaw instances can address one another, exchange end-to-end encrypted A2A envelopes, and route work through approval-aware channels. An A2A local mode keeps peer delivery and authenticated administration reachable while disabling other external gateway and channel surfaces.

First-run onboarding is built around hatching: a new agent asks about the user's work, records useful context, keeps setup links visible in chat, and can send a tailored first-jobs welcome email when an email route is available. The bootstrap lifecycle is recorded as structured onboarding audit events, including messages, welcome mail, workspace file updates, completion, and abort paths.

Credentials stay outside the model context. Secrets live in the encrypted runtime store and SecretRef-backed tools resolve them at execution time, so LLMs see the requested action and approval context, not raw API keys, passwords, or bearer tokens.

Run one local assistant, operate a fleet of role-specific coworkers, or launch HybridClaw on HybridAI Cloud in a few minutes at hybridclaw.io.

Quick Start · Launch Cloud · Installation · Docs · Configuration · Commands · Contributing

Why HybridClaw

Install

Fastest managed launch: HybridClaw on HybridAI Cloud.

Apple Desktop App for macOS:

• Download the signed and notarized Apple Silicon DMG from GitHub Releases.
• Open the DMG, drag HybridClaw.app into /Applications, and launch it.
• The desktop app starts the local gateway and opens the chat, agents, and admin surfaces in a native macOS window.

Linux/macOS CLI one-line installer:

curl -fsSL https://raw.githubusercontent.com/HybridAIOne/hybridclaw/main/scripts/install.sh | bash

Manual npm install:

npm install -g @hybridaione/hybridclaw
hybridclaw onboarding
hybridclaw gateway
hybridclaw tui

Requirements: Node.js >=22.23.4 <23. Docker is recommended for the default sandbox.

First Run

After the gateway starts, open:

For signed macOS desktop builds and future architectures, use the GitHub Releases page.

Desktop wrapper from source:

npm install
npm run desktop

What You Get

Product Strengths

• Validated business skills: production skills use deterministic helpers, fixtures, eval scenarios, and targeted tests so small models can perform useful work through structured actions.
• Guided onboarding: hatching collects useful work context, keeps setup links handy, writes first-job suggestions, and can route the first-run conversation through a stronger onboarding model before returning to the normal default. Its lifecycle is searchable in audit as onboarding events.
• Multi-agent operations: agents can coordinate across local workspaces, hosted HybridAI proxies, and trusted peer HybridClaw instances with A2A pairing, pinned X25519 keys, encrypted envelopes, explicit addressing, inbound messages, reply-back delivery status, admin-visible trust, and an A2A local mode that closes unrelated external surfaces.
• Prompt-level credential isolation: encrypted secrets and SecretRefs keep credential values out of model context while tools receive only the scoped credential material needed at execution time.
• One runtime for many surfaces: web, terminal, Discord, Slack, Teams, email, voice, webhooks, and local API clients all use the same gateway, memory, policy, and audit model.
• Generated apps from chat: describe an app, dashboard, document, game, or tool in web chat; HybridClaw saves the generated HTML artifact into the Apps gallery, refreshes live apps through read-only connector calls, and can share apps through expiring links or Microsoft Teams tabs.
• Secure by default: LLM output is treated as untrusted, risky actions route through approval policy, and every sensitive boundary is visible in audit.
• Model freedom: use HybridAI, the direct OpenAI Responses API, Codex OAuth, other hosted providers, local engines, or named OpenAI-compatible endpoints from the same model picker and config surface. Ordered routing tiers can start routine work cheaply and escalate retry-safe failures while explicit model pins remain authoritative.
• Local Mac models: Labs → Local Models installs and checks a pinned MLX model against the Mac's available memory. Start or stop it from the console, inspect live activity, and select it in chat. Starred tools and skills keep the initial local context compact while catalog discovery makes additional capabilities available during a turn.
• Operator visibility: /admin groups Activity, Agents, Automation, Connectivity, Models, Security, System, and Labs workflows. Page and setting search opens with Cmd/Ctrl+K, settings link to one canonical owner, and legacy admin URLs redirect to their corresponding tab.
• Business-ready extension model: packaged skills, plugins, MCP servers, and SecretRef-backed HTTP tools share the same approval and credential boundaries.