MCP

Let an AI agent generate documents for you.

doctagd exposes a Model Context Protocol (MCP) server. MCP is an open standard, so any compatible agent — Claude, ChatGPT, Gemini, or an agent you build yourself — can discover your saved configs and produce documents from a plain-language request, with no upload and no field-mapping step. It is a thin layer over the same v1 API: your dtgd_ key is the credential, and every quota, drift, and idempotency rule is identical.

Create API key REST API docs

Vendor-neutral

Connect over streamable HTTP at https://doctagd.com/mcp from any MCP client that can send an Authorization header — your API key is the Bearer token.

Self-describing

inspect_config returns a typed schema (date / currency / number / required) so the agent knows exactly what to send and how to format it.

Safe to retry

validate_input is a dry run that costs no quota, and an idempotency key makes generation retries safe.

You still create a config once in the app (upload a template, confirm the mapping). MCP runs configs — it does not author them. Documents and PDF entitlements count against your plan exactly as they do in the app.

Tools

Six tools, each mapping 1:1 to a v1 endpoint. Only generate_documents spends quota.

list_configs

Discover which document templates (saved configs) this account can generate.

inspect_config

Get one config's typed column schema — every column's type, whether it is required, and (for template configs) the placeholder it feeds. Call before generating.

validate_input

Dry-run rows against a config. Returns drift, missing-required columns, and coercion notes. No quota, no document — call it until the input is valid.

generate_documents

Generate one document per row (DOCX, or PDF on Pro/Business). Returns a jobId and a download URL. This is the only tool that uses quota.

check_job

Poll a generation job by id for its status and download URL. Runs are usually synchronous, so the job is often already done.

get_usage

Read the account's remaining monthly document allowance before starting a batch. Reflects shared-license pooling.

Connect

Any MCP client that can send an Authorization header works. Get a key from your account settings (shown once), then point your client at https://doctagd.com/mcp. That includes agent CLIs and SDKs (Claude Code, OpenAI Agents SDK / ChatGPT developer mode, Gemini CLI), IDE assistants (Cursor, VS Code, Cline, Windsurf, Zed), open agents (Goose, Continue), and automation tools with an MCP client node (n8n) — plus anything you build with the official MCP SDKs.

Most MCP clients (config file)

{
  "mcpServers": {
    "doctagd": {
      "type": "http",
      "url": "https://doctagd.com/mcp",
      "headers": { "Authorization": "Bearer dtgd_…" }
    }
  }
}

Claude Code (CLI)

claude mcp add doctagd \
  --transport http \
  https://doctagd.com/mcp \
  --header "Authorization: Bearer dtgd_…"

MCP Inspector (to test)

npx @modelcontextprotocol/inspector

# Transport:  Streamable HTTP
# URL:        https://doctagd.com/mcp
# Header:     Authorization: Bearer dtgd_…

Config-key names vary slightly by client (VS Code uses servers; most others use mcpServers), but the URL and Authorization header are the same everywhere. Discovery manifest: /.well-known/doctagd-manifest.json.

doctagd authenticates with your API key (Bearer), which works with any header-capable client above. The one-click “Add connector” buttons in the consumer Claude, ChatGPT, and Gemini apps use OAuth rather than a header — that flow is on our roadmap, so for now use a client that can send an Authorization header.

A typical agent loop

From a plain-language request to a finished document, with the agent self-correcting before it spends a slot:

The loop

1. list_configs       → find "Service Agreement"
2. inspect_config     → 13 columns, types, which are required
3. (assemble a row from the user's request)
4. validate_input     → catches a missing field / coercion note
5. validate_input     → valid: true
6. generate_documents → jobId + downloadUrl (idempotencyKey
                        keeps retries safe). No human mapping
                        step anywhere in the chain.

The user just asks

“Draft a service agreement for Acme, £12k retainer, starting 1 June.”

The agent assembles + checks

It maps the sentence onto the config's columns and validates — fixing a missing jurisdiction or a date it could not parse — before generating.

You get a file back

A ready .docx (or a .zip for a multi-row batch), fetched from the returned download URL.

Auth, quota & errors

MCP shares the v1 API's auth, limits, and error contract — the tools just forward your key and surface the same responses, so an agent can act on them:

The returned downloadUrl is a time-limited signed link (about 24h) — open it or hand it to the user directly, no header needed. The Authorization: Bearer dtgd_… header also works (e.g. for fixed automations).
A 409 drift error lists missingColumns and suggestions — the agent should remap and retry.
A 402means the monthly document quota is exhausted, or PDF was requested on a Free plan — stop, don't retry.
Requests are rate-limited to 60/min (429).

For key creation, input formats, idempotency, and the full error table, see the REST API docs and plans & limits.