Perplexity Computer alternatives, compared on what decides a self-host call
Not another listicle. Two questions drive every row: can you run it yourself, and who reviews the output. Everything else follows.
| Project | Licence | Runs where | Models | Isolation / review |
|---|---|---|---|---|
| OpenClaw | MIT | Your computer (laptop or shared team server) | Any — Claude, Codex, local models as swappable plugins | Gateway control plane; tools run on host unless sandboxing enabled; channel pairing by default |
| Kortix (Suna) | Source-available, Elastic 2.0 | Your VPC or on-prem; managed cloud optional | Any provider, your own keys | Isolated sandbox per session on its own branch; work lands only through a change request a human approves |
| Eigent | Apache 2.0 | Your infrastructure (Docker + FastAPI + PostgreSQL) | Any / local models | Workspace model; traceability and approval flow |
| Vellum | MIT | Your device (macOS today) | Any | Credentials run in a separate service the model cannot read |
| AnythingLLM | MIT | Your hardware | Any / local | Local-first document retrieval; private vector workspaces |
| Lobehub | Apache 2.0 | Self-hostable | 10+ providers incl. local | Chat workspace; multi-modal, single agent |
| Claude Cowork | Closed | Anthropic cloud | Anthropic only | Apple VM sandbox; per-app permissions |
| ChatGPT Work | Closed | OpenAI cloud | GPT family | Managed; workspace agents |
| Perplexity Computer | Closed | Perplexity cloud | 19 hosted models (Opus 4.6, Gemini, Grok, GPT-5.2) | Isolated Linux sandbox (2 vCPU / 8GB); opaque routing |
Sources: project repositories and documentation (github.com/openclaw/openclaw; github.com/kortix-ai/suna; eigent.ai; vellum.ai; anythingllm.com; lobehub.com), and independent reviews of Perplexity Computer (builder.io, vellum.ai). September 2026.
One honest paragraph each
OpenClaw
An open-source (MIT) AI assistant that runs on your own computer and meets you in the channels you already use — Discord, iMessage, Slack, Teams and more. A local Gateway is the control plane; state, memory and credentials live on your hardware, and models are swappable plugins. There is no paid tier or hosted service. It is the closest thing to a self-owned Perplexity Computer, but it inherits your machine's quirks and needs an always-on box for availability. (github.com/openclaw/openclaw)
Kortix (Suna)
An open-source AI Management System and the self-hostable alternative to Claude Cowork and ChatGPT Work. Your agents, skills, company memory and connectors live in one git repo you own; each session runs in an isolated sandbox on its own branch, and work reaches main only through a change request a human approves. Elastic 2.0 — source-available, not OSI-open: read it, fork it, self-host it, but not resell it as a hosted service. (github.com/kortix-ai/suna)
Eigent
An Apache-2.0 open-source "cowork desktop" with a multi-agent workstation: agents, skills and connectors are packaged once and reused across sessions. Model-agnostic, self-hostable with Docker + FastAPI + PostgreSQL, and shipped with 200+ MCP tools. More setup than a cloud product; more control than any closed one. (eigent.ai)
Vellum
A MIT-licensed personal AI assistant that lives on your device with its own identity and memory. Its distinguishing idea is credential isolation: keys run in a separate Credential Execution Service the model cannot read. Today it is macOS-first; Windows, mobile and web are on the roadmap. (vellum.ai)
AnythingLLM
MIT-licensed, local-first document retrieval and private vector workspaces with agent automation on your own hardware. The strongest fit when the job is "reason over my documents" rather than "operate my desktop end to end." (anythingllm.com)
Lobehub
An Apache-2.0, self-hostable multi-modal chat workspace with a large plugin ecosystem and 10+ model providers including local. A single-agent chat interface rather than a multi-agent orchestration layer. (lobehub.com)
Claude Cowork & ChatGPT Work
Both are closed and cloud-only. Claude Cowork's strength is deep reasoning and a managed Apple-VM sandbox; ChatGPT Work's is its workspace agents inside OpenAI's cloud. Neither can be self-hosted, and both lock you to their lab's models — which is exactly the trade Perplexity Computer asks you to make too. (claude.ai; openai.com)
Perplexity Computer
A closed, cloud-only generalist agent that orchestrates 19 models and runs inside an isolated Linux sandbox with a real browser, filesystem and terminal. Excellent at hands-off, generalist work; impossible to audit, self-host, or run without a Max subscription. (builder.io; perplexity.ai)
How to choose
- Do you need OSI-open, or is source-available enough? If you must redistribute a fork as a hosted service, OpenClaw (MIT) or Eigent (Apache 2.0) fit; Kortix's Elastic 2.0 does not.
- Where does the work run? Desktop/local-first (OpenClaw, Vellum) is simplest. A fleet of org-scale agents that share memory and land work through change requests points to a server-side system (Kortix).
- Who reviews the output? All keep a human in the loop, but the shape differs: Kortix gates everything behind a deny-by-default change request; others use in-app review.