For the way work is done now
Your assistant can send the agreement.
Paste one URL into Claude, ChatGPT or Cursor. The host finds our authorization server, you approve a consent screen, and from then on the assistant works your envelopes inside the scopes you granted — nothing destructive, nothing unlogged.
MCP
One URL. The host does the rest.
AnySign is a streamable-HTTP MCP server with OAuth discovery and self-registration. Paste the endpoint into your assistant; it finds our authorization server, registers itself, and opens a consent screen. Approve once.
- Claude.ai → Settings → Connectors → Add custom connector
- Claude Desktop / Cursor → mcpServers in the config file
- ChatGPT → Settings → Connectors → Create
- Then: “Send the NDA template to dana@…”
// claude_desktop_config.json · .cursor/mcp.json
{
"mcpServers": {
"anysign": { "url": "https://api.anysignhq.com/mcp" }
}
}
// First use: the host reads
// /.well-known/oauth-protected-resource/mcp
// registers itself, and opens the consent screen.
> Which consulting agreements are still waiting on a signer?
→ list_envelopes { "status": "sent", "query": "consulting" }
← { "envelopes": [ { "id": "env_8f14e2",
"title": "Consulting agreement",
"signer_count": 2, "status": "sent" } ] }Fast to connect, hard to misuse.
An assistant is a new kind of user: tireless, literal, and easy to talk into things. The tools it gets and the record it leaves are designed for exactly that.
Consent, not a key
The assistant never holds an API key. It asks for named scopes on a consent screen and gets exactly those, for your workspace only.
Read and send, nothing else
Five tools: list and read envelopes, list templates, draft, send. Voiding, deleting and payments are not tools, so no prompt can reach them.
Attributed in the audit trail
Every action is recorded against the connected assistant, not the person who connected it. Export it from Settings like any other event.
Revoke is immediate
Disconnecting kills the refresh token with the live one. The host is out on its next request, not at the end of the hour.
Any MCP host
Claude, ChatGPT, Cursor, or anything built on the same spec. The host discovers our authorization server and registers itself.
Same allowance, same rules
An assistant draws on the monthly quota the dashboard does, on every plan, and obeys the workspace's IP allowlist and retention policy.
Questions about assistants.
Which assistants work?
Any MCP client that speaks streamable HTTP and OAuth 2.1: Claude (web, desktop and Claude Code), ChatGPT connectors, Cursor, Gemini CLI, Windsurf, and anything built on the same spec. Cursor installs in one click; the host registers itself, so nobody at AnySign has to add it.
What can it do, and what can it not?
List and read envelopes, list templates, create a draft envelope and send it. It cannot void, delete, touch payments or change workspace settings — those are not exposed as tools, so a confused prompt cannot reach them.
How do I know what the assistant did?
Every action lands in the audit trail attributed to the connected app, not to the person who connected it. The same export the Policy tab offers lists exactly which assistant did what, and when.
How do I disconnect one?
From the AI page or Integrations, in one click. Revoking kills the refresh token as well as the live one, so the host is out on its next request rather than at the end of the hour.
Does it cost extra?
No. The MCP server draws on the same monthly allowance as the dashboard and the API, on every plan.