Onyx
Open-source AI chat and search over your internal knowledge, with permissions enforced at query time and full self-hosting.
If your security team has already vetoed hosted AI assistants, Onyx is one of the few serious options that will clear review, because the index, embeddings, and inference can all stay inside your boundary. The retrieval numbers are independently interesting: 72.4 on EnterpriseRAG-Bench against 61.0 for OpenAI File Search is a real gap, and the published $0.69 query receipt is more transparency than most vendors offer. Onyx compares itself directly to Glean, Microsoft Copilot, and Dust. Budget for an engineer to own the rollout, and do not expect Glean's managed polish.
Verified 3d ago · liveness 72/100 · cite: rightaichoice.com/tools/onyx
- Enterprises whose security review blocks hosted AI assistants and need chat over internal knowledge inside their own
- Engineering teams that want to build agents on top of a permission-aware retrieval layer via the MCP server and
- Sales and support teams that need instant cited answers pulled from Slack, Drive, Confluence, and CRM data
- Organizations that want to route different teams to different LLMs, or run open weights on their own GPUs
- Individuals or small teams with no organizational knowledge base to index
- Teams that want zero operational overhead and no self-hosting work
- Buyers who need a working pilot this week without an engineer to own deployment and connector syncs
We scan live Reddit threads, YouTube comments, X posts, G2 reviews and other communities — and hand you an honest verdict in under a minute.
- Honest verdict, not marketing
- Real pros & cons from real users
- Attributed quotes with receipts
3 free scans · no card needed
Skip Onyx if nobody on your team can own Kubernetes or GPU nodes and you want a managed assistant with no infrastructure work at all.
Business is quoted at $20 per user per month on annual billing, so a monthly commitment costs more or is not offered at that rate.
At $20 per user per month on annual billing, Business is priced below Microsoft Copilot and well below Glean, which Onyx explicitly compares itself to. It fits mid-sized teams that already pay for cloud infrastructure and can absorb OpenSearch and GPU costs. Enterprise pricing is custom and adds SSO, on-premise, and region-specific deployment.
In short
Onyx — Open-source AI chat and search over your internal knowledge, with permissions enforced at query time and full self-hosting. Best for Enterprises whose security review blocks hosted AI assistants and need chat over internal knowledge inside their own, Engineering teams that want to build agents on top of a permission-aware retrieval layer via the MCP server and, Sales and support teams that need instant cited answers pulled from Slack, Drive, Confluence, and CRM data. Free to start; paid plans from $20/user/mo.
What's new in Onyx
Checked 3 days agoAcross the latest 5 updates: 5 news mentions.
Onyx studies intrinsic knowledge retrieval scaling challenge
Onyx research examines how corpus scaling affects retrieval recall for semantic search and BM25, relevant to teams weighing index growth against search quality.
Are File Systems All You Need? It Depends...
Onyx tested hybrid search against glob/grep/read file search on 160k-510k enterprise documents and found no single approach wins across scales.
Onyx benchmarks agentic RAG on workplace questions
Onyx reports it outperformed ChatGPT Enterprise, Claude Enterprise, and Notion AI on an internal RAG benchmark of workplace questions.
Lessons from building Onyx Deep Research agents
Onyx details design choices behind its Deep Research agent system, aimed at teams building performant agent pipelines.
Onyx builds internet search workflow to rival OpenAI
Onyx reverse-engineered leading agent web search and rebuilt the two-tool workflow in its product for lower-hallucination internet answers.
What people actually say about Onyx — is it worth it?
We scanned public community sources for Onyx on Jul 3, 2026 and could not establish that the discussion we found is about this tool rather than something else sharing its name. Our own analysis of that scan says the posts were off-subject. Rather than publish a sentiment score built on the wrong subject, we publish nothing here and re-run the scan.
Viability Score
How well maintained and how widely used is Onyx? Built from what the vendor actually publishes (docs, changelog, tutorials, integrations, pricing), whether the site is live, and how much real users discuss it. How we calculate this
Last calculated: October 2026
How we score →Key Features
- Permission-aware enterprise search across 50+ connected sources
- Chat and search UI grounded in internal documents with citations
- Hybrid search combining semantic vectors and BM25 keyword matching
- Advanced RAG with contextual retrieval and LLM-based knowledge graphs
- Custom AI agents with instructions and attached knowledge
- Actions via MCP and OpenAPI tool calling
- Deep research with multi-step agentic search
- Sandboxed code interpreter for data and analysis tasks
- Web search via a two-tool agentic workflow
- Image generation inside the workspace
- Slack integration with a per-channel bot
- MCP server so external agents can call Onyx search as a tool
- Onyx workspace, desktop app, and Chrome extension interfaces
- Granular RBAC with permission inheritance and query-time enforcement
- Self-hosted, VPC, bare metal, and fully air-gapped deployment options
About Onyx
Onyx is an open-source enterprise AI platform that puts a permission-aware chat and search layer over your company's internal knowledge. You ask a question in plain language and Onyx searches your indexed sources in parallel, enforces each source's permissions, and returns an answer with citations. It is aimed at enterprises and mid-sized teams that want grounded answers from their own docs, Slack threads, tickets, and code. The retrieval pipeline is the differentiator. Onyx runs hybrid search combining semantic vectors with BM25 keyword matching, then layers contextual retrieval and LLM-based knowledge graphs on top. On its published EnterpriseRAG-Bench, Onyx reports an overall score of 72.4 on 500 questions, ahead of OpenClaw (68.2), OpenAI File Search (61.0), Amazon Q/Kendra (49.0), and Azure AI Search (48.4). A separate 99-question head-to-head, judged blind by two LLM judges across 220K internal documents, puts Onyx at a 50% win rate against ChatGPT, Claude, and Notion AI. Onyx also publishes cost receipts for a live query: 6.7 seconds and $0.69 total, versus 10.0 seconds and $1.15 for a search-each-app-separately approach. Beyond chat, the product covers custom agents, actions via MCP and OpenAPI, deep research, a sandboxed code interpreter, web search, image generation, and a Slack integration with a per-channel bot. Onyx documents four interfaces: a workspace, a desktop app, an MCP server other agents can call, and a Chrome extension. Connectors span 50+ sources including Slack, Google Drive, Confluence, GitHub, and Salesforce, each synced continuously with the permissions that source enforces. You can route different teams to different models, or run open weights on your own GPUs via vLLM. Deployment is where Onyx separates from managed rivals. The whole stack, including the OpenSearch hybrid index, embeddings, and inference, can run in your VPC, on bare metal, or fully air-gapped, with egress denied. Documents and embeddings stay in your infrastructure and Onyx states your data is never used for training. The Business plan costs $20 per user per month on annual billing.
Behind the Verdict
Onyx is best understood as a retrieval and context layer that happens to ship a chat UI, rather than a chat app that bolts on search. Everything on the site points back to that: the benchmark page, the answer receipts showing retrieval cost ($0.21) and answer cost ($0.48) separately, and the diagram that shows connectors feeding an OpenSearch hybrid index plus Postgres metadata, with vLLM serving inference inside your own account. Strengths. The permission model is the load-bearing feature. Onyx syncs 40+ to 50+ sources and indexes documents with the permissions each source enforces, then re-checks those permissions at query time rather than trusting a stale index. For enterprises where the failure mode is an employee seeing a document they should not, that is the feature that matters. Second, deployment flexibility is genuinely unusual: AWS, GCP, Azure, or your own Kubernetes; single tenancy, VPC, on-premise, or fully air-gapped with egress denied. Third, model choice is real, not decorative. You can swap providers in one click, route each team to a different model, or run open weights on your own GPUs. Fourth, Onyx publishes its own benchmark work, including a study on how corpus scaling affects retrieval recall and a 2026 test of hybrid search versus glob/grep/read file search across 160K-510K documents, which is the kind of thing you publish when retrieval is your actual product. Weaknesses and honest caveats. First, self-hosting is the pitch and it is also the tax. Someone has to run the connector syncs, the OpenSearch index, and the GPU nodes; Onyx mitigates this only at the Enterprise tier with a forward-deployed engineer. Second, benchmark claims are vendor-published. EnterpriseRAG-Bench is Onyx's own open-source benchmark, and the 50% win rate against ChatGPT, Claude, and Notion AI is a tie, not a win, so the honest reading is parity on general workplace questions with an advantage on retrieval-heavy ones. Third, if you live in Microsoft 365 and want Copilot's native integration, Onyx's cross-source connectors are a different shape of product. Fourth, the site's model routing demo names Anthropic and the benchmark names GPT-5.4 as one model tested; the breadth of named model support beyond that is not documented on the pages we could reach. Where it fits. Engineering-led organizations with a DevOps function, a real internal knowledge base (Confluence, Slack, Drive, GitHub), and a security review that has already blocked hosted assistants. Teams that want to build their own agents on top of a permission-aware retrieval layer via the MCP server and developer APIs. Sales and support teams that need cited answers pulled from Slack and CRM without a separate per-app search dance. Where it does not. Solo users and small teams with no corpus to index will get nothing from it. Teams that want to buy a working pilot this week without an engineer to own deployment and connector syncs should look at a managed product instead. And if you
Researching Onyx? Get your full AI stack in 60 seconds.
Free, no signup — tell us your goal and get tools matched to your budget & existing stack.
Real-world workflow fit
Concrete scenarios for the personas Onyx actually fits — and what changes day-one when you adopt it.
You connect Slack, GitHub, Drive, and Confluence, point Onyx at your own VPC with vLLM inference, and ask why the AWS bill jumped $40K. Onyx searches all four sources in parallel, enforces each source's permissions, and names the orphaned perf-test cluster from a specific PR.
Outcome: You find the $1,340-a-day orphaned cluster and tear it down, recovering roughly $40,200 per month, with the answer cited back to the PR and the cost export.
You ask the per-channel Slack bot where a customer's account stands, and Onyx pulls the relevant Slack threads, Drive docs, and CRM records into one cited answer without you leaving Slack.
Outcome: You answer the customer in the same conversation instead of searching three apps, and the answer carries citations your manager can verify.
You register Onyx's MCP server as a tool for your own agent, so when the agent needs company context it issues a single onyx.search tool call with a query and a source list rather than paging through apps one at a time.
Outcome: Your agent gets six passages from three permission-aware sources in one tool call, spending far fewer tokens than a multi-round per-app search.
Use Cases
- Answer employee questions using internal documentation and Slack conversations, with citations back to the source document or thread.
- Diagnose an unexplained infrastructure cost spike by searching GitHub, Drive, Slack, and Confluence in parallel and returning one cited answer.
- Summarize long Slack threads and channels to get up to speed on a topic without reading every message.
- Create custom AI agents for different departments with tailored instructions and attached knowledge.
- Run deep research across internal and web sources for strategic projects.
- Let an external agent call Onyx search as a tool via the MCP server to ground its own answers in company knowledge.
- Ask questions in Slack via the per-channel bot without switching to a separate workspace.
Models Under the Hood
as of 2026-09-24
Limitations
- Self-hosting is the point of Onyx and also its main cost: you run the connector syncs, the OpenSearch hybrid index, and the inference nodes, and air-gapped deployments add more.
- Only the Enterprise tier includes a forward-deployed engineer to help configure and roll out.
- The retrieval numbers are vendor-published on Onyx's own open-source EnterpriseRAG-Bench, and the 50% head-to-head win rate against ChatGPT, Claude, and Notion AI is a tie rather than a win.
- GPT-5.4 is the only model named in the benchmark materials we could reach, so the full breadth of model support is not documented on the pages we saw.
- Business pricing is $20 per user per month on annual billing.
as of 2026-10-04
Verification history
We have re-verified Onyx 9 times since . Each pass re-reads the vendor's own pages and re-checks every listed field against that evidence; passes where nothing had changed are marked as such.
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
Showing the 6 most recent of 9 verification passes.
Free to cite with attribution — this page re-verifies continuously.
12-month cost
Project the real annual outlay, including the implied monthly cost when only an annual tier is published.
Vendor list price only. Add-on usage, seat overages, and contract minimums are surfaced under Hidden costs & gotchas.
Plans compared
For each published Onyx tier: who it actually fits, and what it adds vs. the previous tier. Cross-reference the cost calculator above for projected annual outlay.
Free Trial
$0
Ideal for
A team evaluating Onyx before committing, with at least one cloud source to connect and an engineer available to run the trial seriously.
What this tier adds
Starting tier and free entry point: a trial of the Business plan with the chat and search UI and access to all major LLMs, at no cost.
Business
$20/user/mo (annual billing)
Ideal for
Mid-sized teams and enterprises that need cited answers over internal knowledge and are comfortable running the stack themselves or in their VPC.
What this tier adds
Adds the full production feature set over the trial: custom AI agents, actions via MCP/OpenAPI, 40+ app connectors, web search, deep research, code interpreter, image generation, the Slack integration, and developer APIs, at $20 per user per month on annual billing.
Enterprise
Contact us
Ideal for
Large organizations with security, compliance, or data-residency requirements that a Business plan cannot satisfy, including air-gapped and on-premise deployments.
What this tier adds
Adds OIDC/SAML SSO, on-premise and region-specific deployments, white-labelling, custom integrations, hook extensions, data exports, invoice billing, volume discounts, dedicated support, and an Enterprise SLA on top of everything in Business.
Where the pricing makes sense
The company stage and team size where Onyx's pricing actually pencils out — and where peers do it cheaper.
At $20 per user per month on annual billing, Business is priced below Microsoft Copilot and well below Glean, which Onyx explicitly compares itself to. It fits mid-sized teams that already pay for cloud infrastructure and can absorb OpenSearch and GPU costs. Enterprise pricing is custom and adds SSO, on-premise, and region-specific deployment.
Setup time & first value
How long it actually takes to get something useful out of Onyx — broken out by persona, not the marketing-page minute.
A Business trial gives you a working workspace in minutes if you connect a cloud source like Slack or Drive. A self-hosted or VPC deployment realistically takes days to weeks: someone has to stand up OpenSearch, Postgres, and GPU inference nodes, configure the connectors, and verify permission sync. Enterprises get a forward-deployed engineer to shorten that.
Switching to or from Onyx
How to bring data in from common predecessors and how to get it back out — written for the switcher, not the buyer.
- →From ChatGPT Enterprise or Claude Enterprise: point Onyx at the same knowledge sources and route teams to the same underlying models, keeping the answers and citations inside your VPC.
- →From Glean: replace the managed index with Onyx's self-hosted OpenSearch hybrid index and keep the same Slack-facing chat workflow.
- →From Open WebUI or LibreChat: keep the open-source chat habit but move from general model access to permission-aware retrieval over your own corpus.
- →From per-app search: retire the search-Drive-then-search-Slack-then-search-Confluence loop in favour of one parallel query.
- →From Microsoft Copilot: move cross-source retrieval off the Microsoft 365 boundary onto infrastructure you control.
- ↗To Glean: if you want a fully managed index and are willing to give up self-hosting and air-gapped deployment.
- ↗To Microsoft Copilot: if your organization has standardized on Microsoft 365 and native integration outweighs cross-source retrieval.
- ↗To ChatGPT Enterprise: if you want a general assistant with web knowledge and do not need permission-aware indexing of internal sources.
- ↗To Dust: if you want a managed assistant-building layer with different deployment trade-offs.
Integrations
Resources & Guides
Tutorials & Learning
YouTube returned 6 videos for “Onyx”, and we withheld 6: 6 could not be judged, because “Onyx” is a single word that other videos use for other things. We are showing none, because we could not prove any of them are about Onyx.
Official links
Featured Head-to-Head Comparisons
Onyx vs Presto Voice
Choose Presto Voice if you operate a QSR drive-thru and want to automate ordering with proven upselling (up to 6% revenue lift). Choose Onyx if you need a secure, open-source AI assistant grounded in your enterprise documents and data. These tools serve entirely different use cases: one for physical voice ordering, the other for knowledge management and research.
Onyx vs Screenplayiq
ScreenplayIQ wins for screenwriters needing market-oriented script analysis with predictive box office stats; Onyx dominates for enterprise teams wanting a customizable, data-sovereign AI search/chat platform grounded in internal knowledge. Your choice depends on whether your priority is script monetization or enterprise knowledge management.
Onyx vs Truleo
Truleo is purpose-built for law enforcement intelligence needs, excelling at connecting siloed data like jail calls and body cameras. Onyx offers a flexible, open-source platform for enterprise knowledge retrieval and deep research, with recent benchmarks showing superior RAG performance. Choose Truleo if you're in law enforcement; choose Onyx for general enterprise AI grounded in internal documents.
Popular in Enterprise Search & Internal Knowledge
Frequently Asked Questions
Categories
Used Onyx? Help shape our editorial sentiment research.