Docs / Self-hosted AI workspace: bring your own model

Article · Self-hosting & models

Self-hosted AI workspace: bring your own model

For product teams that need to control their own data and models, Maskin is fully self-hostable, Apache-2.0, and bring-your-own-model. Self-hosting is the structural answer for teams searching for an open source alternative to Claude Cowork, Codex, or ChatGPT workspaces — the capability without the vendor’s lock-in.

Why self-host

Three buyers ask for self-hosting: data sovereignty (the work never leaves your infrastructure), model freedom (you choose the models, including local ones), and auditability (the code is open and inspectable). Hosted vendor workspaces cannot answer these — your data and your model choice are the product.

Open-core self-host is also a governance stance: the closed loop runs on objects you own, inside a repository you control, under a permissive license.

Apache 2.0: inspect every line

Apache 2.0 is a permissive open-source license — use it, modify it, run it commercially, keep your changes private or contribute them back. For an AI workspace, that license matters more than usual: the tool decides which agents see what, and you should be able to read how.

Bring your own model

The workspace is model-agnostic at the protocol layer because it is MCP-native. You decide which model runs your agents: Claude for reasoning-heavy shaping, Codex or a local model for constrained tasks, or a fine-tuned model for your domain. See how model providers and keys work.

Sandboxed agents

Agents execute in sandboxed sessions — isolated environments that cannot touch your private keys or production state outside the workspace’s authority. Autonomous operation is safe because execution is contained, and every step is inspectable. See the security model for auth, data isolation, and key handling.

Managed vs self-hosted

If you want the same loops without running infrastructure, book a hosted trial. If the work must never leave your infrastructure, self-host.

Pricing

• Self-host — always free.
• Pro — $20/seat/month, agent credits included.
• Team — $200/workspace/month, unlimited seats.
• Enterprise — your own LLM and infrastructure.

An open-source alternative to Claude Cowork / Codex

The “open source alternative to X” searches are filling fast with desktop coding apps. What is missing is a closed-loop workspace for product work: signals in, bets shaped, tasks executed, outcomes validated — self-hostable and model-agnostic. The open source alternative to Claude Cowork page compares that shape against the vendor and open-source options directly. That is the niche Maskin occupies, and it sits in the composable camp of the agent-workspace market rather than the AI-employees one (the two camps compared). For how the same object-model argument plays against the crowded coding-agent lane specifically, see the self-hosted AI agent control plane comparison — twelve OSS control planes shipped in 2026 that all stack as kanbans of coding tasks.

For a wider read on where the self-host-persona OSS agent field sits in 2026 — stateful-agent frameworks, workflow-node builders, RAG chat frontends, coding-agent control planes — the field survey of self-hosted AI agent repos is the honest category map, and names which shape fits which shape of work.

Start with the get-started guide or the self-hosted setup guide to stand up an instance on your own hardware.

FAQ

How hard is self-hosting?

Clone the repo, follow the onboarding guide, and connect your model provider — agents run in sandboxed sessions out of the box.

Can I switch between self-hosted and managed later?

Yes — the license and object model are yours; hosted and self-hosted are the same product.

Is self-host truly free?

Yes — the open-source software is Apache 2.0; you pay for infrastructure only.