MCP quickstart — on-brand documents for AI agents

Most “AI + PowerPoint” tools stop at generating a file. DocMark’s MCP server gives your agent something different: a brand-faithful, verified document arm. The agent writes the substance; DocMark applies your corporate brand deterministically, runs its verification gates, and hands back either the finished artifact or the findings to fix — so your agent closes the loop itself.

Availability. Available now, and included in every plan — MCP is a thin client of the same /v1 API and draws on the same monthly render credits. Signed-in customers download the server from the portal’s API tokens page; questions to hello@docmark.md. Also runs self-managed — on-prem or your own cloud tenant.

1. Get the server and its runtime

Two things before the config below will work:

• docmark_mcp.py — download it from the portal’s API tokens page (app.docmark.md → API tokens → Connect your AI assistant), which also has a one-file Claude Desktop extension. Workspace admins create the API token on the same page; everyone else asks an admin or us. It is a single file; put it anywhere and note the absolute path.

• Python 3.10+ on your PATH, plus the MCP runtime:

  pip install "mcp<2"
  

  Pin the major version. A plain pip install mcp installs 2.0.0, and the server fails to start. Claude Desktop bundles Node, not Python — you need your own.

2. Connect the server

DocMark runs a standard MCP stdio server. Add it to your MCP client’s config — for Claude Desktop / Claude Code:

{
  "mcpServers": {
    "docmark": {
      "command": "python",
      "args": ["/absolute/path/to/docmark_mcp.py"],
      "env": { "DOCMARK_API_TOKEN": "dmk_xxxxxxxx.xxxxxxxx" }
    }
  }
}

On macOS, use python3 rather than python — stock macOS has no python on PATH.

The server is a thin client of the /v1 API — the token is the same tenant-scoped bearer token, kept control-side (never in the model’s context).

3. The tools your agent gets

• list_themes — the brands/looks available to your tenant.
• generate_deck / generate_document — submit a document job (the tool docstring teaches the strict spec + an exemplar, so the agent gets the shape right the first time).
• job_status — poll a job; a failed job returns its findings.
• fetch_result — download the verified artifact to a path.

4. The loop that makes it reliable

agent authors content  ->  generate_deck  ->  job_status
                                                 |
                         done? -> fetch_result   |  failed? -> findings
                                                 |               |
                                                 <---- fix & resubmit

When a verification gate withholds the deliverable, the agent receives why (the findings) and resubmits a corrected job — no human in the loop for the mechanical fixes. That’s the same render-then-verify discipline behind every DocMark output, exposed as a tool.

Why route your agent through DocMark

• Brand-faithful, not brand-ish — your (or your customer’s) real corporate colours, fonts, and template, applied deterministically. Per-customer branding is built in.
• Verified before delivery — the gate withholds slides that are blank or carry nothing at all, plus structurally broken files, and your agent gets the findings to fix and resubmit.
• Multi-format — not just slides: Word, Excel, brochures, newsletters, and narrated video from the same door. (Most MCP document tools are PPTX-only.)
• Contract-versioned — the underlying API is pinned (1.18.0), so your agent integration is stable.

Next: full API reference · deployment modes · request access