Skip to content

AI Integration (MCP)

TongueToQuill exposes a Model Context Protocol (MCP) endpoint. Any MCP-aware AI client — Claude, OpenAI clients, custom agents — can use it to produce a TongueToQuill document from natural language and return a link the user can open.

The endpoint takes no API key. Each document is bound to your account afterward by a short claim handshake (account pairing) — the AI hands you a link, you open it and sign in, and the draft is yours. Some deployments add a six-digit device code to that step. See Account Pairing. For the invited beta deployment, see Beta Setup.

What it can do

  • Generate USAF / USSF memorandums and other supported document types from a description.
  • Build from a specific published template when you give it the template's short code.
  • Revise the draft it just made, on request, before you save it.
  • Retrieve the format schema for a Quill so the AI can produce valid frontmatter.
  • List the Quills available on the current deployment.

What it can't do

  • Read or edit documents already saved in your library.
  • Browse or search the template gallery, or use your starred templates (it can load a specific template only from a short code you give it).

Each conversation produces one draft the AI can revise; once you save it, further editing happens in the web app.

Setup (Claude)

  1. Open Claude → Settings → Connectors.
  2. Click Add Connector.
  3. Enter the URL: https://tonguetoquill.app/mcp.
  4. Save.

Once connected, Claude can call the document tools — list_quills, get_template, get_specs, create_document, and update_document, plus claim_document where the deployment uses device-code claiming. For how a document is bound to your account, see Account Pairing.

If your organization runs its own TongueToQuill deployment, replace the host with your own (e.g., https://airmark.example.mil/mcp). The endpoint is only available when MCP is enabled on the deployment.

Setup (other clients)

Point any MCP client SDK that speaks Streamable HTTP at https://<your-host>/mcp. SSE is not used; the call sequence and claim handshake match the Claude setup above.

The standard call sequence:

  1. list_quills — discover which document types this deployment supports.
  2. get_specs — retrieve the format rules and field blueprint for the chosen Quill.
  3. Edit the blueprint into the document.
  4. create_document — submit the markdown. Receive a claim token and a claim link.
  5. Hand the user the link and ask them to open it and sign in. That claims the draft; it's now theirs to edit, copy, or export.

Where the deployment uses device-code claiming, step 5 also asks the user for the six-digit code the link shows, and a final claim_document call replays the claim token and that code to bind the document.

Starting from a template? If the user gives you a template short code (from the gallery's Copy for AI chat action), call get_template first, then get_specs for the format it returns and continue from step 3.

To revise after the claim, call get_specs then update_document with the same claim token — no new link or code needed.

For advice on what to put in the prompt that drives this sequence, see Effective Prompting.

Example

User prompt to Claude:

Draft a memo to 20 FW/CC about the upcoming safety inspection on 14 May. From 20 OG/CC. Mention mandatory attendance for all flight commanders. Sign it Lt Col Smith, Commander.

Claude calls get_specs for usaf_memo, composes the markdown, calls create_document, and hands you the claim link.

You open the link and sign in. The claimed memo is yours — review, edit, save a copy to your library, export as PDF.

When to use this vs. the editor directly

Use the editor when Use the MCP server when
You know exactly what you want to write You want to dictate a memo conversationally
You're starting from a template You don't remember which template you need
The document is sensitive enough you don't want it touching a third-party AI You're happy for an AI to draft an initial version
You need to attach files, indorsements, or set specific fields the AI doesn't know You have a routine memo with a clear, stateable purpose

Whatever the AI generates lands in the editor. You always get the final review.

Next