holaboss-ai/holaOS

HolaOS is a local-first Electron desktop workspace that puts agent and application side by side on the same screen — not as a chat transcript feeding into a separate tool, but as a live UI surface driven by the agent while you watch and intervene. The README calls this "HolaApps": any app from the in-workspace marketplace opens as an interactive panel next to the agent, context flowing both ways. Click something yourself and the agent already knows; let the agent drive and you see every move.

The model story is deliberately pluralist. Built-in options include cost-efficient models like Kimi K3 and GLM 5.2 for routine volume, and heavier options like GPT 5.6, Claude Opus 5, and Fable 5 for harder tasks — all on one account, no API keys required by default. Bring your own keys for OpenAI, Anthropic, or any compatible endpoint and those calls route through your own account instead. The agent layer is similarly non-prescriptive: Claude Code, Codex, and holaOS's own agent run in the same workspace over shared memory, tools, and skills. You pick per task without rebuilding context.

Memory is stored as plain files on your own disk — readable, editable, deletable. The README is explicit that no copy of your work sits on their servers, and frames this not as a policy promise but as an architectural fact: the data simply never travels there. Every tool connection and file access requires explicit approval.

The integration surface is wide: over 50 one-click OAuth connectors, MCP server support, scheduled automation, and output as real .xlsx, .pptx, and .docx files rather than chat text. Chat platforms — Slack, Feishu, DingTalk, WeChat — connect so the agent reads actual conversation threads rather than your summary of them.

HolaHub, the recipe-sharing layer, lets users publish and install complete agent configurations — model, prompt, skills, MCP, integrations — in a single click. The framing is that you see how a result was made, not just the result itself.

The license is a modified Apache 2.0 with additional commercial-distribution and branding conditions, which matters if you're evaluating it for internal redistribution. Installation is a single curl command on macOS, Linux, or WSL, and the README includes a verbatim prompt you can hand to a coding agent to run the setup on your behalf — a small but telling detail about who the intended user is.

The honest caveat: this is a README, so the gap between described capability and shipped behavior is unknowable from here. The model names listed (GPT 5.6, Claude Opus 5, Fable 5) are marketing-tier labels that don't map cleanly to publicly documented model versions, which warrants some skepticism about the built-in model claims specifically.