What Is Onyx (onyx.app)?

Onyx is an open-source AI enterprise search platform that connects to your company's knowledge sources and returns accurate, fast, token-efficient context to both people and AI agents. It's built for teams that want AI to be their primary interface to internal information — searching GitHub, Google Drive, Slack, Confluence, and similar systems in parallel rather than one at a time. If your problem is "our AI agents can't find the right internal answer, or they burn too many tokens and too much time looking," Onyx is aimed squarely at that.

What problem it solves

Most AI search setups query one source at a time. The page's own side-by-side example makes the failure mode concrete: asked "Why did our AWS bill jump $40K this month?", a traditional MCP-style agent ran search_drive("cost export") (9 files), search_slack("aws bill") (0 results), search_confluence("cloud costs") (74 pages), and search_github("perf-test") (rate limited) — then needed 16 more rounds, one source at a time, until the context window filled up with only 3 of 10 relevant docs. The agent concluded the cost growth was "expected" and missed the actual cause.

Onyx's approach is to search all indexed sources in parallel, enforce permissions, and return a small set of relevant documents. In the same example it surfaced 10 relevant docs and identified the culprit: a perf-test-eu cluster from PR #2213, spun up for a July load test and never torn down, costing $1,340/day since July 18 — recoverable at $40,200/month.

How it works, in plain terms

• Connects to enterprise knowledge across sources like GitHub, Drive, Slack, and Confluence.
• Searches in parallel instead of sequentially, so it doesn't exhaust the context window before finding the answer.
• Enforces permissions, so results respect who is allowed to see what.
• Returns context to humans and agents — the same retrieval layer serves both.

The page describes the search pipeline as sending your question "out in several forms" to improve matching, and claims it outperformed industry leaders on an open-source benchmark.

Speed and cost

Onyx publishes a worked comparison on its own site:

The headline claim is "faster answers at lower cost" — in this example, roughly 3.3 seconds faster and about $0.46 cheaper per query. Treat these as vendor-published example figures, not guaranteed results for your workload.

Benchmark claim

Onyx reports results on EnterpriseRAG-Bench (500 questions, overall):

• Onyx GPT-5.4: 72.4
• OpenClaw: 68.2
• OpenAI File Search: 61.0
• Amazon Q (Kendra): (figure truncated in the source)

The benchmark is described as open source, so you can inspect the methodology rather than take the ranking on faith.

Who it's for

The page cites adoption signals of 1M+ queries per week and 1,000+ teams, and says "the world's top teams trust Onyx to be their primary interface to AI." That points to organizations that:

• Have knowledge spread across many tools (code, docs, chat, wikis) and need one retrieval layer over all of them.
• Run AI agents that need grounded, permission-aware context without blowing the context window.
• Want an open-source option rather than a fully closed search product.

What to check before deciding

• Pricing and plan limits — the site links to a Pricing page and offers "Try for Free" and "Book a Demo," but the excerpt doesn't state actual prices or what the free tier includes. Check the Pricing page directly.
• Connector coverage — confirm the specific sources you rely on are supported, since the examples shown are GitHub, Drive, Slack, and Confluence.
• Permission model — verify it maps to your identity provider and access rules, since "permissions enforced" is a core claim.
• Benchmark fit — EnterpriseRAG-Bench is a general benchmark; your corpus and query patterns may behave differently.

If you mainly need a single-source search or don't run AI agents over internal data, the parallel-retrieval advantage matters less. If your agents are missing answers or costing too much per query across many sources, that's the case Onyx is built for.