holaboss-ai/holaOS

The core bet holaOS makes is that agent lock-in is the real problem, not capability. Rather than building yet another agent with its own isolated context, it builds a shared substrate — memory, tools, integrations, app surfaces — that any agent can drive. Claude Code, Codex, or the built-in holaOS agent all operate over the same workspace. Switch between them mid-project and nothing resets.

The memory design is worth examining closely. Context and project history are stored locally as plain files you can open and edit directly. That's a meaningful architectural choice: it means the memory layer isn't a proprietary blob or a cloud service you're dependent on, it's just files. The README claims this memory is also structured and embedded so retrieval is actually useful rather than just archival — though there's no detail here on how that embedding works or what the retrieval mechanism looks like under the hood.

The HolaApps concept is the most distinctive piece. Instead of agents that report results back as chat text, apps open as live interactive surfaces alongside the agent. The Notion example in the README is illustrative: the agent navigates inside the actual Notion interface while you watch, and you can take over at any point. This is a different interaction model than most agent UIs, which treat the app as a black box and the chat window as the only surface. Whether it holds up in practice depends entirely on how well the browser automation and app embedding actually work — the README doesn't go into implementation depth there.

The model roster listed — Kimi K3, GLM 5.2, GPT 5.6, Claude Opus 5, Fable 5 — reads like marketing copy for models that may or may not exist at those version numbers at any given time. The bring-your-own-keys path for OpenAI, Anthropic, or any compatible endpoint is the more durable value proposition for anyone who already has provider relationships.

On the integration side, the claim is 50-plus one-click OAuth connections covering Gmail, Notion, Slack, GitHub, Linear, and others, plus MCP server support. The MCP compatibility matters: it means the tool surface isn't closed, and community-built servers can extend what any agent can reach.

The license is modified Apache 2.0 with additional commercial-distribution and branding conditions — not fully permissive, so read the LICENSE file before building anything commercial on top of it. The Electron desktop runs on macOS (both Apple Silicon and Intel), Windows, and Linux, and the install path is a single curl command or a five-step manual flow pinned to Node 24.14.1.

The honest question this project leaves open is whether the shared-memory, multi-agent-workspace model is genuinely useful or whether most people just pick one agent and stay there. If you're already context-switching between Claude Code and Codex on the same project, holaOS has a real answer. If you're not, the added surface area may not justify the setup.