Signvoy

傳送文件以供電子簽名、追蹤簽署狀態，並下載完成的PDF。OAuth 2.1搭配動態客戶端註冊——無需API金鑰。

文件

Connect AI agents to Signvoy

Signvoy exposes a Model Context Protocol server at POST /mcp. Claude, Cursor, and any other MCP-compatible client can send documents for signature, check signing status, and retrieve signed PDFs — without leaving the conversation.

No API key

16 tools

Streamable HTTP

OAuth 2.1 + DCR

Why there is no API key

Most e-signature MCP servers ask you to generate a credential and paste it into a config file. Some also require you to register an application and copy your AI client's OAuth redirect URIs into it before anything works.

Signvoy implements OAuth 2.1 with Dynamic Client Registration, so your client does that work itself. You paste a URL; the client discovers the authorization server, registers, and opens a browser for you to consent. Nothing long-lived is ever written to a config file on disk.

Why this matters: a static API key in claude_desktop_config.json is a plaintext secret with no expiry, sitting in a file that AI agents can often read. An OAuth grant is scoped to one client, revocable on its own, and carries no secret you could leak by sharing your config.

Connect in one step

Pick your client. There is no single block that works everywhere — clients disagree on the top-level key, the URL key, and whether a transport field is required, and Claude Desktop does not read remote servers from a config file at all.

Cursor

Claude Desktop

Claude Code

VS Code

Windsurf

Gemini CLI

Other

One entry in mcp.json. Nothing to install.

Add to Cursor

Manual setup

1 Open Customize → MCPs in the Cursor sidebar, or edit the config file directly.
Add Signvoy to your global config so it is available in every project. Use.cursor/mcp.json in a project folder instead if you only want it there.
```
{
  "mcpServers": {
    "signvoy": {
      "url": "https://api.signvoy.com/mcp"
    }
  }
}
```
3 Restart Cursor, then click the server's authenticate control. A browser opens for you to sign in and approve the connection. Full Cursor guide

Whichever you use, your client opens a browser to authenticate the first time it connects. Nothing long-lived is written to disk.

How auth works

You do not need to understand this to use the server — your client handles all five steps. It is documented because knowing the flow makes failures easy to read.

1Your MCP client sends a request to POST /mcp with no token.
2Signvoy replies 401 with a WWW-Authenticate header naming the discovery document.
3The client fetches /.well-known/oauth-protected-resource/mcp and registers itself through Dynamic Client Registration. No human step.
4A browser opens for you to sign in to Signvoy and consent. One time only.
5The client stores the access token and retries. Every later request is automatic.

What agents can do

16 outcome-shaped tools, all prefixed signvoy_ to avoid collisions with other MCP servers. Full parameters and return shapes are on the tool reference.

Read · 8 tools

signvoy_list_workspaces read

List every workspace the authenticated user belongs to.

signvoy_list_templates read

List templates in a workspace. Defaults to active (non-archived).

signvoy_get_template read

Get one template in full, including its recipient role keys.

signvoy_list_documents read

List documents newest-first, with status filter and text search.

signvoy_get_document_status read

Per-recipient signing state for one document.

signvoy_get_audit_trail read

Append-only event log for chain-of-custody verification.

signvoy_get_signed_pdf_url read

Presigned download URL for the certified signed PDF.

signvoy_list_contacts read

Search the workspace address book to turn a name into an email.

Write · 8 tools

signvoy_create_document write

Create a draft document from a published template.

signvoy_update_document writeidempotent

Fix a draft before sending — recipients, name, message, or routing mode.

signvoy_send_document writeidempotent

Email signing requests to every recipient.

signvoy_remind_document write

Send one reminder email to all pending recipients.

signvoy_remind_recipient write

Send one reminder email to a single pending recipient.

signvoy_create_signing_link writeidempotent

Embedded signing URL for one pending recipient, no login required.

signvoy_void_document destructive

Irreversibly void a document and revoke all signing links.

signvoy_create_contact write

Save someone to the workspace address book ahead of time.

Workspace selection

An OAuth token identifies a user, not a workspace, so every tool accepts a workspace_id and your assistant decides which workspace each call acts on. Omitting it works only when your account belongs to exactly one workspace. Past that, Signvoy will not guess: the call fails and names the workspaces you could have meant, so the assistant can ask you and retry.

If you belong to more than one workspace, ask your assistant to “list my Signvoy workspaces” before starting a signing workflow. A document is created inside one workspace and cannot be moved afterwards, so it is worth settling before any draft exists.

Safety model

Sending a contract is irreversible in a way that most agent actions are not. Four properties of the server are designed around that.

Create and send are separate tools — signvoy_create_document only produces a draft and hands back a review_url. Nothing is emailed until a second, explicit signvoy_send_document call. An agent cannot conflate the two into a single step.

Destructive actions are annotated — signvoy_void_document carries the MCP destructiveHint, so well-behaved clients prompt before calling it.

Every action is attributed — mutating calls write an audit event tagged actorSource: "mcp" with your user ID, IP address, and user agent. Agent activity is distinguishable from dashboard activity in signvoy_get_audit_trail.

One revocable grant per client — each client that connects gets its own grant, which lets it act as you. Revoking one leaves the others working, and there is no static secret to rotate.

Human in the loop: real emails reach real counterparties. Confirm the document name and recipient list before allowing signvoy_send_document, and be deliberate about running an agent fully unattended — a prompt-injected document could otherwise influence what gets sent.

Manage grants under Workspace Settings → AI Agents, where each connected client shows when it was last used and a button to revoke it. The same operations are available over the API as GET /v1/mcp-grants and DELETE /v1/mcp-grants/:id.

FAQ

Do I need an API key to use the Signvoy MCP server?

No. Signvoy uses OAuth 2.1 with Dynamic Client Registration, so your MCP client registers itself and obtains a token through a browser consent screen. You paste one URL into your client config — there is no key to create, no client secret to generate, and no redirect URI to configure.

Which AI clients can connect to Signvoy?

Any client that speaks MCP over streamable HTTP, including Claude Desktop, Claude Code, and Cursor. Signvoy implements the standard OAuth discovery documents, so clients that support remote MCP servers work without a Signvoy-specific plugin.

Can an AI agent send a contract without my approval?

Creating and sending are separate tools. signvoy_create_document only produces a draft and returns a review_url for a human to open; a second explicit call to signvoy_send_document is what actually emails recipients. The bundled signvoy_send_for_signature prompt instructs the model never to skip that review step.

How many tools does the Signvoy MCP server expose?

16 tools: 8 read tools for listing templates and documents, checking per-recipient status, reading the audit trail, and downloading signed PDFs, plus 8 write tools for creating, updating, sending, reminding, generating embedded signing links, and voiding.

Is agent activity recorded in the audit trail?

Yes. Every mutating tool call writes an audit event tagged actorSource: "mcp" along with your user ID, IP address, and user agent, so actions taken by an AI agent are distinguishable from actions taken in the dashboard.

Which workspace does an agent act on?

Whichever one the tool call names. An OAuth token identifies a user, not a workspace, so every tool accepts a workspace_id. You can omit it only if your account has exactly one workspace; otherwise the call fails and lists the candidates, so the assistant can ask you which you meant and retry.