Running it on your own hardware
Many buyers ask one question first: does the text of a prompt leave our network? This page shows what you can run on your own servers, how MitDesk decides that something is "local", and what MitDesk cannot promise.
Who this is for: administrators and security owners who run models on their own servers.
You need: Edition: Local and self-hosted models, endpoints and the safety checks in the first sections are part of Basic. Sovereign Mode, local model operations and the certification gate need Business. See Editions. A local AI server can also be agreed as part of Ultimate.
Two ways to use your own model server
| Way | What you set up |
|---|---|
| Ollama | Add a model with the provider Ollama. It needs no key. Its address is set in the settings (block AI keys, field Ollama) and defaults to http://localhost:11434. The Python package ollama must be installed for the Python that runs Odoo. MitDesk names the package if it is missing. |
| Any OpenAI-compatible server | For example vLLM or llama.cpp. Create an AI endpoint in Configuration ▸ Engine ▸ AI endpoints. Give it a name and a base URL that ends in /v1. Then add a model with the provider Named endpoint and choose that endpoint. A key is optional, because many self-hosted servers have none. |
A database can have many endpoints. Each has its own address and its own key. You can run a server in your office and also use an EU-hosted service, and one does not overwrite the other.
The base URL is required on purpose. MitDesk never falls back to a vendor's servers for an endpoint.
Use Test connection on the endpoint form. It checks that something answers at the address and shows the hosting class MitDesk assigns. It does not pin versions or detect changes.
How MitDesk decides "on-premises"
The hosting class (On-premises, EU cloud, Non-EU cloud) is read-only. You say which server to use. MitDesk decides what it counts as.
- Ollama always counts as on-premises. MitDesk does not check the address. Point Ollama only at a machine you control.
- For an endpoint, an address is on-premises when it points inside your own network: a loopback, private or link-local IP address,
localhost, or a name ending in.local,.internal,.lan,.localdomainor.home.arpa. - A short built-in list of EU hosts counts as EU cloud.
- Everything else, and anything MitDesk cannot read, counts as Non-EU cloud.
MitDesk reads the address as text and never looks it up. It checks again at every call and does not trust the stored value.
A public name is never "local"
Imagine your GPU server is in your office, but a rented machine in a data center ends the TLS connection and passes the traffic on. The model really runs on your hardware. MitDesk still does not call it on-premises, because the text of each prompt exists for a moment on a machine you do not own. For the on-premises class, the address MitDesk connects to must be inside your own network.
Where a key may be sent: the endpoint allowlist
The hosting class answers "is this inside our network?". A second check answers "may MitDesk send a key to this host?".
The risk is an address that was changed without the settings screen, for example in a database row or a restored backup. The next call would then send your key to whoever owns that host. So MitDesk checks the address at the moment it attaches the key.
- Loopback addresses (
localhost,127.0.0.1) are always allowed. The key does not leave the machine. - These hosts are allowed:
api.openai.com,api.anthropic.com,generativelanguage.googleapis.com,api.mistral.ai,api.groq.com,openrouter.aiandapi.deepseek.com. - Any other host, including a private address such as
10.0.0.5, must be added once to the system parameterminddesk.endpoint_allowlist. Separate hosts with commas or new lines.
If a host is not allowed, MitDesk refuses and names the host and the parameter.
Keys are encrypted
Keys you enter in MitDesk are stored encrypted. The encryption uses two secrets together: the master password (admin_passwd) in odoo.conf, and the database.secret of the database. A copy of the database alone does not give the keys. If admin_passwd is still the default admin, the protection is weak and MitDesk logs a warning until you change it. An endpoint's key field is write-only.
Two exceptions:
- A key from an environment variable (
MINDDESK_<PROVIDER>_API_KEY) is read directly and is protected by your host. - A key saved by an older release stays plain text until it is saved again. After an upgrade, save each key once.
Guards that work together
- The cloud deny list. In the settings (block Cloud egress, field Blocked from cloud egress) you list Odoo models whose data may never go to a cloud provider. Only a local provider skips it. EU and non-EU cloud must pass it. A block raises an error and writes an audit row that stays even if the action is rolled back. This guard is part of Basic.
- Data-class routing. Decides for each turn by the kind of data. See Sovereign routing. Business.
- Sovereign Mode. One switch that limits every AI turn to local models, and optionally EU models. Business.
The deny list and routing also cover knowledge-search embeddings and outbound MCP tools and webhooks. Sovereign Mode does not, because they are not AI turns. For a closed setup, keep the embedding model local, and do not add outbound MCP servers or webhooks that you have not checked. See Knowledge and sources.
Scanned documents
MitDesk reads a scanned file on your own server. No OCR engine is included, because which one is acceptable is your decision. Install pytesseract (with Tesseract), easyocr or paddleocr on the server, and pypdfium2 for scanned PDFs. Without an engine, MitDesk says that it could not read the scan. It does not send the scan elsewhere. As an alternative, choose a Document reader model in the settings. Only a vision model on an on-premises or EU-hosted endpoint is offered. See Document intelligence.
Operate the local server
With the Sovereign module, the settings block Sovereign Mode has two buttons for Ollama:
- Check local endpoint shows whether the server answers and its version.
- Sync local models records what the server really serves: whether each model is available, its digest and its real context size.
On an Ollama model, use Pin served version to pin the digest you reviewed. If the server later serves a different digest under the same name, MitDesk marks the model as drifted. With the Compliance module, drift also creates an integrity incident for the compliance responsible.
For a local model, an administrator sets a tool-calling reliability tier (Certified, Limited, Unrated) by hand. MitDesk never sets it. A local model must be certified for the job role before a digital employee can run on it. See Hiring a digital employee.
The monthly AI budget counts only paid cloud providers. Local providers are never counted or blocked by it. This includes a Named endpoint and the OpenAI-compatible provider, even when the address is a paid EU service. Use the vendor's own limits for such a service.
In chat, the conversation shows the hosting of the last request, for example Local — nothing left your server.
What the attestation does not claim
The Compliance module can export an egress attestation. See Compliance support. It reports the guard setup, the guard events that happened, and where the recorded model calls went. It states that it does not guarantee zero egress.
What MitDesk cannot prove
If a prompt goes to a third party, MitDesk can tell you that it went there, to which hosting class, and by which decision. It cannot tell you what the third party did with it. No software on your side can prove that a provider did not keep a prompt. That is a matter of contracts. For this reason the processing record links each provider's data-processing agreement.
One setup avoids the question for the turn itself: the model answers from an address inside your own network. Even then, an outbound MCP tool, a webhook and the embedding provider send data on their own account. A closed setup needs a local model, a local embedding provider, and no outbound MCP server or webhook that you have not checked.
If it does not work
If a key is refused because of the host, see Troubleshooting.
Related
- Sovereign routing: how one turn is decided.
- Connecting a model: create the model and endpoint.
- Knowledge and sources: where you choose the embedding provider.
- Connecting an external AI: the MCP side, including outbound tools.
- Compliance support: the attestation and the processing record.
- The modules: what each module adds.
- Troubleshooting: common refusals.