# holaOS Bets Your Agent Is the Disposable Part > Source: > Published: 2026-08-18 13:09:01+00:00 [AI](https://sourcefeed.dev/c/ai)Article # holaOS Bets Your Agent Is the Disposable Part The trending workspace makes memory, tools, and OAuth grants outlive Claude Code and Codex sessions. [Priya Nair](https://sourcefeed.dev/u/priya_nair) The pitch behind [holaOS](https://github.com/holaboss-ai/holaOS), currently climbing GitHub's trending charts at roughly 9,700 stars, is easy to miss under the "All in One AI agent workspace" branding. The interesting claim isn't that it runs agents — everything runs agents now. It's that the agent is the *swappable* part. You run [Claude Code](https://claude.com/claude-code), [Codex](https://openai.com/codex), or holaOS's built-in agent inside one local workspace, and they all share the same memory, tools, integrations, and files. Close the app, come back a week later, switch agents entirely — the context persists, stored (per the README) locally as plain files you can read and edit. That's a bet about where value accrues in the agent stack, and it's worth taking seriously even if you never install this particular Electron app. ## The agent loop is commoditizing; the environment isn't Two years ago the hard part was the agent itself — the tool-calling loop, the planning, the recovery from failure. That's now table stakes. Claude Code, Codex CLI, Block's [Goose](https://block.github.io/goose/), and a dozen open-source clones converge on essentially the same architecture, and [MCP](https://modelcontextprotocol.io) standardized how all of them reach tools. What none of them solved is continuity. Every serious agent user has felt this: the session ends, the context dies, and tomorrow's session starts by re-explaining the project. The ecosystem has been attacking that gap from several directions. The `CLAUDE.md` /`AGENTS.md` convention made plain-text files the de facto portable memory format. Letta (née MemGPT) turned memory management into a service. Mem0 shipped a shared memory layer that multiple agents reach over MCP. holaOS's move is to fuse all of that into a *place*: a desktop environment where the workspace — not the agent — owns the memory, the OAuth grants to 100+ services (Gmail, Slack, Notion, Linear, GitHub), the MCP servers, the browser automation, and the scheduled recurring jobs. Under the hood it's an Electron shell over a Fastify runtime with a SQLite state store, plus vector recall via sqlite-vec; agents plug in through a swappable harness that gets handed a compiled context scoped to the workspace. If you currently maintain separate MCP configs for Claude Code, Cursor, and Codex — duplicating the same server definitions and re-authorizing the same services in each — you already understand the problem holaOS is monetizing. Configure once, and any harnessed agent inherits the lot. ## What adoption actually looks like Installation is a one-line curl-to-bash that sets up a Node monorepo, so budget the usual skepticism plus a read of the script first. Coverage from earlier this year described it as macOS-only with Windows and Linux "in progress," so check current platform support before committing a team to it. Models are either built-in (frontier models via holaOS's own access) or BYOK against OpenAI, Anthropic, and compatible endpoints — the built-in path implies an account relationship with holaboss, which matters for anyone doing this on a corporate machine. The realistic adoption unit isn't "replace your terminal." It's the recurring, long-horizon jobs that CLI agents handle badly precisely because they're stateless between runs: weekly competitor sweeps, feedback triage, report generation into real `.xlsx` and `.docx` outputs, agents reachable from Slack or Telegram. That's the workload where durable memory plus scheduling genuinely replaces a pile of cron scripts and re-pasted context. ## The trust ledger doesn't balance yet Here's where my enthusiasm cools. A workspace that holds OAuth tokens for a hundred services *and* hands autonomous agents a browser is the largest blast radius you can construct on a laptop. That demands unusual transparency, and holaOS currently offers less than usual. The license is a modified Apache 2.0: internal use is fine, but hosting it for third parties or embedding it in a sold product requires a commercial license from holaboss, and you can't strip the logo from the frontend. Fair enough — that's the standard open-core defensive crouch. Less fair: independent coverage from April 2026 described the project as MIT-licensed, and the repository's visible history now begins at a single squashed "Initial commit" dated late July — months *after* that coverage, when the project already had nearly 2,000 stars. Whether that's a routine pre-launch cleanup or a quiet relicense, the audit trail a security team would want is gone. A repo showing 9,700 stars alongside three open issues is either miraculously polished or not yet stress-tested by real users; my money's on the latter. There's also a gap between marketing and mechanics. "Memory as plain files you can edit" is the pitch; the architecture also keeps continuity data in a machine-owned `.holaboss/` directory and a SQLite runtime database. That's a sensible design — but your escape hatch, if you ever leave, is only as good as the plain-file portion. ## The pattern beats the product My read: holaOS is directionally right and prematurely trusted. The direction — agent-agnostic environments where memory, tool auth, and durable state outlive any one vendor's agent — is where this space is clearly heading, because it inverts the lock-in. If your context lives in the workspace, Anthropic and OpenAI compete for you every session. But you can adopt the pattern today without adopting the product. Keep project memory in plain markdown that every agent reads (`AGENTS.md` already does this across tools). Centralize MCP server definitions instead of per-tool configs. Treat your agent CLI as a stateless executor pointed at durable, inspectable state you own. That's the actual insight holaOS is packaging — and it's yours for free, under whatever license you like. Watch the repo; wait on the OAuth grants until the project's history is longer than its star count is tall. ## Sources & further reading - [holaboss-ai/holaOS](https://github.com/holaboss-ai/holaOS)— github.com - [holaOS: Agent Environment for Long-Horizon Work](https://dudarik.com/en/blog/holaos-mcp/)— dudarik.com - [holaOS Architecture Overview](https://deepwiki.com/holaboss-ai/holaOS)— deepwiki.com - [holaOS LICENSE (modified Apache 2.0)](https://raw.githubusercontent.com/holaboss-ai/holaOS/main/LICENSE)— raw.githubusercontent.com [Priya Nair](https://sourcefeed.dev/u/priya_nair)· AI & Developer Experience Writer Priya covers AI frameworks, developer productivity tooling, and the startup ecosystem across South and Southeast Asia, bringing a researcher's rigour and a practitioner's empathy to every story. She is deeply sceptical of benchmarks and asks hard questions so her readers don't have to. ## Discussion 0 No comments yet Be the first to weigh in.