Deployment
Run Agnes your way
Every company gets its own Agnes instance, so your tables and documents never sit in an index shared with anyone else, and you can have it run for you or run it yourself in your own cloud account. You bring your own AI key, and model calls go only to the provider you chose, which reads what it needs to answer.
Facts checked 8 November 2026.
Two ways to run it
Choose who runs the instance
Both paths give you the same thing: a single-tenant instance that serves your company only. The difference is who keeps it running.
- Managed: we run your dedicated instance. We run a separate instance for you, with automated setup on Google Cloud. It includes daily disk snapshots kept for 30 days and automatic upgrades, so your IT team does not patch anything. This is the faster start.
- Self-run: your IT runs it in your cloud account. Your IT runs Agnes with Docker Compose on a virtual machine in your own cloud account, including AWS or Azure. You own operations, backups and upgrades, and no account of ours is involved. This is the path when policy says the stack must live in your account.
We'll walk you through how running it yourself is packaged and supported on a call.
Your AI key
Bring your own AI key
- Anthropic with your key. Chat runs on Anthropic's API under your own key. The model relationship is yours.
- Claude in your own Google Cloud project. Or run Claude through Vertex AI in your own Google Cloud project and region, so the calls sit under your existing Google Cloud controls.
- Chat on the self-run path. On the self-run path chat is off until you configure it. A documented setting can run chat sessions in containers on the same host.
What leaves the instance
A short list of what leaves, and where it goes
- Model calls go to your chosen provider. The provider has to read a question to answer it. Your organisation decides which provider that is.
- Claude Code is its own route. Analysts who work in Claude Code reach Claude Code's model provider directly. Ask us for a plain-English walkthrough of that route.
- Some tables can be copied. In local mode, tables are copied onto the Agnes server, and the command-line tool can copy the ones a person may see to a laptop. A server-only setting keeps a table off the laptop. In remote mode the table stays in its source and only results travel.
So a deployment stays inside a boundary you care about only when the model provider, and every copy of the data, are inside that same boundary. We would rather show you the list than promise more.
Side by side
The two paths, compared
| Managed: we run it | Self-run: your IT runs it | |
|---|---|---|
| Who keeps it running | We operate your dedicated instance | Your IT owns operations, backups and upgrades |
| Where it runs | Automated setup on Google Cloud | A virtual machine in your cloud account, including AWS or Azure |
| Upgrades and backups | Automatic upgrades and daily disk snapshots kept for 30 days | Manual upgrades; you set up backups and monitoring |
| AI provider | Anthropic with your key, or Claude in your Google Cloud project | The same, once you configure chat |
| Who shares the index | Nobody: single-tenant | Nobody: it is your own instance |
For your IT team
What to ask on the demo
- Which path fits our policy, and how is it packaged and supported for us?
- Who can reach the server and its storage, and what is recorded about their actions on the host?
- How do we get the container image? It is moving to private access, so you will need an authorised token or your own mirror.
- Which sign-in method will we use, and what does the analyst command-line tool need to reach on the network?
- Which tables should stay server-only, so they never land on a laptop?
Where another product fits
If you want a vendor to operate it inside your cloud
Glean offers Customer Hosted: a vendor-operated service in your Google Cloud or Amazon Web Services account that you do not patch yourself. Glean says it is not a traditional self-hosted model, and it keeps support access. If you want exactly that arrangement and have it in writing today, Glean is worth a look. Microsoft 365 Copilot and ChatGPT Enterprise are vendor-hosted products with their own residency options.
Where your data goes
You choose the AI provider
Agnes sends model calls to the provider your deployment is configured to use: Anthropic's API, or Claude running in your own Google Cloud project through Vertex AI. We don't claim your data never leaves your cloud: the AI provider has to read a question to answer it. What we can tell you is exactly where each call goes, and that your tables and documents sit in your own single-tenant Agnes instance, not in an index shared with other companies.
If people use Agnes from Claude Code, Claude Code talks to its own model provider as usual, and the tables a person is allowed to see can be copied to their laptop for fast local analysis.
Sources
Sources behind this page
Agnes facts on this page come from the Agnes product documentation, checked on 8 November 2026. Facts about other products come from their public pages linked below, read the same day. Products change; if something here is out of date, tell us on the demo.
- About Glean deployment (opens in a new tab) Glean-operated Customer Hosted, as the packaged comparison.
- Glean Cloud Prem (opens in a new tab) Last updated 29 September 2026.
Proof
Tested on real questions, graded by the customer
Cuesta Partners asked the same 10 business questions three times each and graded every answer themselves. Agnes passed 100% of graded answers. ChatGPT with SharePoint passed 78%, and Claude with SharePoint passed 70%. One firm, one set of questions: directional, not a general ranking.
Questions
Questions this page can answer
Each answer stays inside the sources behind this page. A gap in the sources is stated as a gap.
Will our documents sit in an index shared with other companies?
No. Each company gets its own single-tenant Agnes instance. Your tables and documents live there and in no shared index.
Can our own IT run it in our cloud account?
Yes. Your IT can run Agnes with Docker Compose on a virtual machine in your own cloud account, including AWS or Azure, and owns operations, backups and upgrades. We'll walk you through how running it yourself is packaged and supported on a call.
What actually leaves the instance?
Model calls. When someone asks a question, the provider you configured has to read it to answer. That is Anthropic's API with your key, or Claude in your own Google Cloud project through Vertex AI. If people use Claude Code, it talks to its own model provider, and tables a person may see can be copied to their laptop.
Whose AI key does Agnes use?
Yours. Chat runs on Anthropic's API with your key, or on Claude inside your own Google Cloud project, so the contract and the billing for the model stay between you and that provider.
How do our people sign in on the self-run path?
With Entra, Okta or another OpenID Connect provider, Google Workspace, a Keboola stack, local passwords or magic links. LDAP and SAML are not supported.
Who at the operator can reach our instance?
Whoever runs the server can reach the server and its storage, so choose the operator on purpose: your own IT, or us for a managed instance. An optional daily export of operating numbers is off unless it is enabled, and an admin can disable it.
Bring one question to a meeting.
These pages stay inside what the product docs and the linked vendor pages supported on 8 November 2026.