Mode 2 — Client cloud

The promise

Everything on-prem does, plus AI — running inside your own cloud tenant. DocMark is deployed in your Azure, AWS, or GCP subscription and managed by your team. Your content stays inside your cloud boundary; it never enters DocMark’s infrastructure. AI authoring, drafting, chart derivation, agentic generation, and optional web research with cited sources all run there, on the LLM subscription you bring.

We will never call this mode “zero-egress” — because it isn’t. Content sent to your chosen model provider leaves DocMark’s process and reaches that provider through one governed channel, under your account and your contract. We think telling you that plainly is a feature.

Bring your own LLM

You bring the AI relationship; we publish the guidance. DocMark’s own authoring runs under your own account with the LLM provider you bring. With the AI engines other than the agentic one it is one governed call with a capped output-token budget (in API mode); the optional agentic engine instead runs a multi-step agent session, bounded by a turn limit and a time limit but with no per-job spend cap, so set a spend limit with your provider. The provider is Anthropic (Claude) today, with OpenAI, Azure OpenAI, and Google Gemini on the roadmap. (Prefer another provider now? Your own agent can author with any LLM and call DocMark’s /v1 API or MCP server to render — with every engine except the optional agentic one, the branded render itself uses no LLM.) The key, the spend caps, the data-processing terms, and the provider clearance are all yours.

How it works

1. Your team submits a topic, a brief, or existing content — through the web app, the /v1 REST API, or your own AI assistant over MCP. In this mode, you run the /v1 API and the MCP server yourself, inside your tenant.
2. DocMark makes one governed call with a capped output-token budget (in API mode) under your own account (BYO key) to the LLM provider you bring (the optional agentic engine runs a multi-step agent session instead, with no per-job spend cap) — Anthropic (Claude) today; other providers on the roadmap. Your security team clears the provider directly; you set the spend caps; by default the provider does not use API data for training, and you confirm that in your own contract.
3. The model authors the outline or researches the topic — web research returns cited sources that ship in the deliverable.
4. Still inside your tenant: branded render, then the verification gate — on by default, and turning it off takes a second, deliberate override as well, named in your handover documentation (see security). With every engine except the optional agentic one, rendering is deterministic, tested code (pixels never from AI). With the agentic engine, an AI agent writes and runs the build script and is instructed to run its own visual review; DocMark does not verify that it ran, so the deck gate is applied to its output as well. Pass and it’s delivered (a deck sent in through the watched folder also carries its contract-protected-api label); fail and it’s withheld with findings.

The BYO-LLM story, as a feature

• BYO subscription (recommended): the AI relationship is yours, on your paper, cleared by your security team — we never see your content’s API traffic or its terms.
• Your cloud boundary: content and rendering stay inside your tenant. DocMark is software you run, not a service your content is shipped to.
• Fail-closed attestation (watched-folder intake): the watched-folder intake will not start an AI engine unless the operator has attested that the data-processing agreement is signed. Web-app, /v1 and MCP jobs are not checked this way: they run on your key, under your contract with your provider.
• Machine-written labelling: the watched-folder intake writes the egress class on every deck it renders, and a deck that touched a model can never carry a zero-egress label. Web-app, /v1 and MCP jobs are not labelled.

Who it’s for

Teams whose content is confidential, but whose security organisation can accept running AI inside their own cloud tenant under their own provider contract — and who want the productivity of AI authoring, cited research, and a /v1 API plus MCP server they operate themselves, without giving up local rendering, in-boundary verification, and audit-grade provenance.