Comparison
Lyzr Open CoWork unites chat, agents, knowledge and governance in one portable workspace.
Open WebUI, LibreChat and Onyx entries reflect each project's public documentation as of September 2026.
Compare Lyzr Open CoWork with
| Capability | Lyzr Open CoWork | Open WebUI | LibreChat | Onyx |
|---|---|---|---|---|
| Everyday AI workspace | ||||
| Models and routing | Frontier, open-weight, local and fine-tuned models; a router picks the cheapest capable model per request and fails over between providers | Ollama and any OpenAI-compatible provider; user picks the model | Major providers and OpenAI-compatible endpoints; user picks the model | Major and self-hosted providers; user picks the model |
| Chat and skills | Files, web search, deep research, voice, images, memory, projects, skills and sharing; imports ChatGPT and Claude history | Files, web search, voice, images, memory, skills, shared chats | Files, web search, voice, images, memory, projects, skills, ChatGPT import | Files, web search, deep research, voice, images, projects, sharing |
| Documents and artifacts | Word, PowerPoint, Excel, PDF and HTML in a side panel, in the company design system; pages publish to an internal link | HTML, SVG and JavaScript visuals beside the chat | React, HTML and Mermaid artifacts; Office files through code execution | Craft builds decks, reports, spreadsheets and web apps in a sandbox |
| Desktop and browser | macOS and Windows app that works on local files and commands with approval; browser automation in the employee's own session | Desktop app in early alpha; Open Terminal runs commands | No official desktop app or built-in browser automation; code workspaces are highly experimental | Desktop app for access and notifications; Craft browser runs in a sandbox |
| Knowledge and data | ||||
| Knowledge and retrieval | Knowledge bases with RAG (retrieval-augmented generation), agentic RAG and graph RAG over a company knowledge graph | Knowledge bases with hybrid search, reranking and agentic retrieval | Vector search over uploaded files through a separate RAG API; no agentic RAG or graph RAG | Agentic RAG over a hybrid index with AI-generated knowledge graphs |
| Semantic data modeling | Semantic models map business terms to databases and warehouses, so agents answer accurately from structured data | Not documented | Not in LibreChat; ClickHouse announced a semantic layer for ClickHouse data only in November 2025 | Not documented |
| Connectors and tools | Drive, SharePoint, Slack, Jira, Confluence, Salesforce, mail and calendar; any MCP (Model Context Protocol) server or tool, with approval before writes | Drive, OneDrive and SharePoint pickers; MCP and OpenAPI tools with approval prompts | MCP servers and API actions | 50+ synced connectors; permission sync in paid editions; MCP actions |
| Agents | ||||
| Building agents | Agents and multi-step workflows built by describing them in plain language, then tested, versioned and scheduled | Model presets with prompt, tools and knowledge; scheduled automations | No-code agent builder and subagents; scheduled chats are labelled experimental | Custom agents with knowledge and actions; Craft scheduled tasks |
| Agent registry and marketplace | Central registry and marketplace; the assistant finds the right agent for a task, or builds one when none exists | Models shared by user or group | Marketplace of agents shared by users of the same deployment, by category, with viewer, editor or owner access | Agents shared by user or group |
| Existing and external agents | Imports and runs agents from any framework, such as Microsoft or LangChain agents, through A2A (agent-to-agent protocol) or HTTP | No A2A support; users have proposed it | No A2A support; outside agents enter only as chat models through an OpenAI-compatible adapter | Not documented |
| Marketplace governance | Personal agents run freely; team and company agents pass IT approval with change-request evidence before they are published | Access grants per user or group | Access lists per user or group | Group permissions in Enterprise Edition |
| Agent control plane | Deployment pipelines, versions, observability and kill switches for agents on any framework, deployed to Kubernetes or major cloud agent runtimes | Manages only its own model presets; no deployment pipelines | Manages only its own agents, including through a beta management API; no deployment pipelines | Manages only its own agents; no deployment pipelines |
| Testing and improvement | Simulation before release, shadow tests on historical data, drift measurement and a continuous agent improvement engine | Model arena and leaderboards | Trace export to Langfuse | Trace export to observability tools |
| Governance and cost | ||||
| Identity and access | Single sign-on, directory sync, admin and member roles, team model allow lists, central keys and per-team feature switches | Single sign-on, directory sync, groups and per-model access in the free edition | Single sign-on, directory login, Entra group sync, per-role settings | Single sign-on, directory sync, roles and model access by group in paid editions |
| Tokenomics and budgets | Cost attributed to every person, team, agent and model; budgets and limits in tokens or money on any consumption, with hard stops | Third-party add-ons; built-in analytics estimate cost | Per-user token credits with auto-refill | Rate limits globally, per user or per group |
| Observability and audit | Trace of every request and agent step with tokens and cost; audit trail of every action and admin change; exportable dashboards | Token and cost analytics, audit logging, OpenTelemetry | Token dashboard without money cost; OpenTelemetry and Langfuse | Usage analytics in paid editions; audit event stream to security tools |
| Responsible AI | Policies on personal data, secrets and banned topics that warn, redact or block; sensitive requests stay on private models | Custom Python filters and pipelines | Automated moderation | Custom code hooks to strip personal data or reject queries, paid |
| Deployment and ownership | ||||
| Where it runs | Hosted by Lyzr, in the customer's cloud or on premises, including an air-gapped appliance with no outbound connection by default | Docker, Kubernetes, pip; air-gapped possible | Docker and Helm; local or cloud | Docker, Kubernetes, Terraform; air-gappable; Onyx Cloud |
| Open source and portability | Open source, modular and portable; agents are plain git repositories in an open format, with zero vendor lock-in | Source-available; BSD-3 plus a branding clause above 50 users | MIT licence; owned by ClickHouse since November 2025 | MIT core; single sign-on, groups and analytics under an enterprise licence |
Talk to us.
Connect employees to organizational knowledge, approved AI models, and specialized digital coworkers — then safely turn answers into action.
Enterprise-grade security and compliance.
Open Cowork is built and operated against the certifications enterprise security reviews require.
GDPR
SOC 2 Type II
ISO 27001
HIPAA
CCPA