Onyx AI is an AI assistant that lets you search and chat in natural language across the documents and business apps scattered around your organization. It connects to services such as Google Drive, Slack, Confluence, and Salesforce, so a question like “What changed in last month’s invoicing flow?” is answered using your own internal data as the source. Its defining trait is that the core is published as open source under the MIT license, so you can stand it up on your own servers or any cloud you like. If you would rather not run the infrastructure yourself, Onyx also offers a managed cloud version.
Where a general-purpose chat AI answers from what the model learned during training, Onyx answers from what your company already knows. That makes it a fit for teams whose internal wiki search never surfaces the right document, or whose information is split across a different tool for every department.
Key Features
- RAG search across internal data: It combines hybrid search (keyword plus vector), contextual retrieval, and LLM-based knowledge graphs to generate answers grounded in your internal documents. Answers cite their sources, which makes verification straightforward
- A large connector library: Connectors for major business apps — Google Drive, Slack, Confluence, Salesforce, Notion, GitHub, and more — ship out of the box (the official site says 40-plus; the GitHub README says 50-plus). They support scheduled syncing and access control that inherits the permissions from the source tool
- Deep research, web search, and code execution: Beyond internal data, it can also use web search (Serper, Google PSE, Brave, SearXNG, and others), run multi-step deep research, execute code in a sandbox, and generate images
- Custom AI agents and MCP integration: You can create agents with their own instructions and reference knowledge. Actions via the Model Context Protocol (MCP) and OpenAPI are supported, so external tool calls can be built into a workflow
- Bring your own LLM: Alongside commercial models from OpenAI, Anthropic, and Google, you can point it at a locally hosted model — including configurations that keep sensitive data away from external APIs
- Self-hosted or cloud: Docker, Kubernetes, Helm, and Terraform are supported, and a one-line install script gets you started. There is a lightweight Lite setup that runs in under 1GB of memory and a Standard setup that includes indexing and background processing
Pricing
| Plan | Price | What’s included |
|---|---|---|
| Self-hosted (Community Edition) | $0 (MIT license) | Core features run free on your own servers. Infrastructure costs and operations are on you |
| Business (cloud) | $20 per user/month (billed annually) | Chat and search UI, access to major LLMs, custom agents, MCP/OpenAPI actions, 40-plus connectors, web search and deep research, code interpreter and image generation, Slack integration, developer APIs, Google OAuth and RBAC, community support. Free trial available |
| Enterprise | Contact for pricing | Everything in Business plus OIDC/SAML SSO, on-premise and region-specific deployments, early access to new features, white-labeling and custom integrations, data exports, invoice billing, volume discounts, and dedicated support with an SLA |
Pricing reflects information as of August 2026. Check the official site for the latest details.
Note that enterprise-oriented capabilities such as SSO, audit logs, and advanced analytics are distributed as a separately licensed Enterprise Edition, distinct from the MIT-licensed core.
Pros & Cons
✅ Pros
- The core is open source (MIT), so you can run it in your own environment with no license fee
- You can configure it so data never leaves your infrastructure, which suits organizations handling confidential information
- Connectors for the major business apps are already there, so you can search across information that stays where it lives
- The LLM is swappable, letting you choose based on cost, performance, or data protection policy
- A cloud version exists, so you can start managed and move to self-hosting once your operations are ready
⚠️ Cons
- Self-hosting is free of license cost, but server bills, updates, and backups become your responsibility
- The full setup runs vector indexing and background workers, so it needs a reasonable amount of machine resources
- Results depend heavily on how carefully connectors and permissions are configured, and the initial setup takes real effort
- Features often required for enterprise rollouts, such as SSO and audit logs, sit in the paid edition
- It is somewhat heavy for a single individual; a general-purpose chat AI is often simpler for solo use
Comparison with Similar Services
| Criteria | Onyx AI | Glean | General chat AI (ChatGPT, Claude) | Dify |
|---|---|---|---|---|
| Primary use | Cross-company data search and AI assistant | Enterprise search and assistant | General conversation and content creation | Platform for building LLM apps |
| Self-hosting | Yes (MIT core) | No (SaaS) | No (SaaS) | Yes (open-source edition) |
| Internal data integration | 40-plus connectors built in | Many connectors built in | Connector features exist but limited in scope | You build it yourself |
| LLM choice | Commercial or local, your call | Managed by the vendor | Fixed to the vendor’s own models | Free choice |
| Entry price | Free self-hosting, or $20/user/month | Contact for pricing | Individual monthly plans available | Free self-hosting available |
Who Is It For
- Teams whose information is scattered across an internal wiki, Slack, and cloud drives, where “I don’t know where it’s written” is the normal state
- Organizations that do not want to hand confidential data to an external SaaS, or cannot do so for regulatory reasons
- IT departments that can run their own servers and want to launch internal AI search while keeping license costs down
- Development teams that want to switch LLMs depending on the task or the budget
- Companies that want to validate the value on the cloud version first and move to self-hosting later
Summary
The value of Onyx AI lies in being able to run an AI assistant that answers from your own information assets, as open source, on infrastructure you control. The connectors and RAG implementation are solid, but the payoff depends on data connections and permission design, so plan for a meaningful amount of setup work. If you have the operational resources, a practical path is to stand up the self-hosted edition in a test environment, confirm how well it fits your own data, and then consider the cloud or Enterprise editions when you roll it out company-wide.