Skip to content

Private deployment and private AI ​

Private deployment is a proposed optional service. Some clients ask: "Can DutyGraph run in our own environment, on a private AI model?" The answer is yes, in steps. This page lists four tiers, what each one gives the client, what is built today and what each tier still needs.

WARNING

Planned, optional. Much of the groundwork is built: the app runs on any server with plain PostgreSQL, and it has connections for Azure OpenAI, AWS Bedrock and Google Gemini, plus Microsoft, Google and AWS options for email, speech, file storage and keys. None of those cloud connections has been run against a live account yet. A connection to a model the client hosts, an installer, single sign-on and offline use are not built. The Status box at the end says which part is which.

Who it is for ​

ClientWhat they usually mean
A company with an AI policy"Our data must go only to the AI we already have a contract with (Microsoft, AWS or Google)." This is tier 1.
A company with a strict security team"Nothing runs outside our own cloud account." This is tier 2.
Healthcare, finance or legal with no outside AI allowed"The AI model must run on our servers." This is tier 3.
Defense and some regulated sites"No internet at all." This is tier 4.

Ask "What does your security team require?" before quoting. Most clients who ask mean tier 1 or 2.

The four tiers ​

TierWhat the client getsWhere the AI runs
1. Hosted, their AI contractHosted DutyGraph. AI calls go to the client's own Azure OpenAI, AWS Bedrock or Google Gemini account.The client's cloud AI account, under their contract
2. Their cloud accountDutyGraph installed in the client's Azure, AWS or Google account, with their database and their email.The client's cloud AI account
3. Their own modelTier 2, plus an open model the client hosts (for example Llama, Qwen or Mistral served by vLLM or Ollama).The client's servers
4. Fully offlineTier 3 with no internet connection.The client's servers

What each tier needs ​

Tier 1: hosted, their AI contract ​

  • Built: the Azure OpenAI, Bedrock (Claude) and Gemini connections. The client's key is saved encrypted for their company only.
  • Needs: a live test with a real account for that vendor (text, streaming, reading files). Today the AI vendor is chosen for the whole server, so choosing it per company is a small extra change.

Tier 2: their cloud account ​

  • Built: the app runs on any server with Node and plain PostgreSQL, and runs its own background jobs. Email through Microsoft 365, Gmail or Amazon SES. Keys in Azure Key Vault, AWS KMS or Google KMS. Files in Amazon S3 or Azure Blob storage.
  • Needs: a packaged installer (app, database and scheduled jobs), an install and upgrade guide for the client's IT team, a first live install, and single sign-on (Entra ID or Okta), which most large companies require.
  • Note: employees' Claude or ChatGPT reach DutyGraph's MCP server over the internet, so the client must allow that one address in.

Tier 3: their own model ​

  • Needs: a connection to any OpenAI-compatible model server on a private address, and quality testing of the Copilot, Pitch to Package, the North Star writer and Task Card fill on the client's chosen model. Server-side transcription is optional: voice answers on private links already transcribe in the person's browser.
  • Note: answers are weaker than Claude or GPT unless the client runs a large model. Say so up front.

Tier 4: fully offline ​

  • Needs: everything in tier 3, plus serving the browser transcription model from DutyGraph itself, an email connection to the client's own mail server (SMTP), and offline release bundles.
  • Does not work offline: company web research, phone push notifications, and employees using claude.ai or ChatGPT with the MCP server. Inside an offline site, employees use the DutyGraph web app.

How it would be sold ​

It is an add-on on top of the standard mapping fee and retainer. The client pays their own cloud and hardware bills. Tiers 3 and 4 start with a short paid scoping that confirms the model, the hardware and the security rules, credited toward the install.

TierFirst build, our timeLater clientsDraft one-time feeDraft yearly fee
1. Hosted, their AI contractAbout 1 week1 to 2 days$10kNone, or +10%
2. Their cloud accountAbout 5 weeks1 to 2 weeks$40k to $75k+$50k
3. Their own modelTier 2 plus 3 to 5 weeks2 to 3 weeks$100k to $150k+$75k to $100k
4. Fully offlineTier 3 plus 2 to 3 weeks3 to 4 weeks$175k to $250k or more+$150k

The fees are a first draft (November 14, 2026), not a price list.

INFO

Status. Built: the app runs off Vercel on plain PostgreSQL; the switchable connections for AI (Azure OpenAI, Bedrock, Gemini), email (Microsoft 365, Gmail, SES), speech (Azure, Amazon, Google), files (S3, Azure Blob) and keys (Key Vault, KMS), chosen with DUTYGRAPH_PROFILE; voice transcription in the browser. Built but not tested against a live account: every one of those cloud connections. Not built: a connection to a self-hosted model, an installer, single sign-on, an SMTP email connection and offline bundles. Not decided: the final prices. Technical detail: service ports and deployment profiles.

Source: gitbook/product/roadmap/private-deployment.md in the DutyGraph repository.

DutyGraph. Pages label each feature Built, In review or Planned.