Skip to main content
Abigail runs a Model Context Protocol server. Point an MCP capable assistant at one URL and it gains 45 patent prosecution tools: docket, deadlines, office action analysis, the guided response wizard, USPTO forms, and agent workflows.

​
What you get

The same product as abigail.app, reached through your assistant instead of a browser. The server is a stateless proxy in front of the identical backend the web app calls, so the account, the docket, the tenant isolation, and the credit balance are one and the same. A case added through an assistant appears on My Docket; a draft generated in the web wizard downloads through MCP. There is no sync step and no second copy of the data. Every tool result is JSON, which matters if you are not using a chat client at all. The protocol is the integration surface, so an independent application can speak MCP directly and parse structured results rather than scraping prose.

​
The lifecycle

Sign up and funding happen on the web. Everything after that happens wherever your assistant lives.

​
Signing in

Every tool requires a signed in account. A request that carries no valid Bearer token is answered with HTTP 401 and a WWW-Authenticate header naming the protected resource metadata document, which is what a compliant client needs to discover the rest on its own. Your assistant performs this on your behalf. There is no key to copy, no pre issued secret, and no onboarding email: the server supports dynamic client registration, and the only code challenge method is S256.
The MCP server does not accept abi_sk_ API keys. It uses OAuth 2.1. See Authentication for the comparison.

​
Your platform

Setup, per client

Claude, ChatGPT, Cursor, and custom clients, with the discovery chain spelled out step by step.

Match the web app

Every Abigail screen reproduced as a tool sequence, plus the known gaps.

Tool reference

All 45 tools, what each one requires, and the web app feature behind it.

Billing and limits

What is deducted, when, and the current rates.
In short: Claude takes a custom connector URL under Settings, Connectors. ChatGPT adds an MCP server the same way, and reads the tool annotations the server advertises so it can tell a lookup from a write. Cursor and other clients take a mcpServers entry pointing at the URL. A custom client speaks streamable HTTP directly.

​
Billing

MCP usage draws on the same credit balance as the web app, priced identically. AI usage is deducted per token, and document exports carry a fixed fee. Call check_credits at any point to read the balance, and see Billing and rate limits for the current rates. Purchases happen on the web only, at abigail.app/settings?tab=billing. Tools that report a balance work normally inside an assistant; topping up does not.

​
For independent developers

The transport is streamable HTTP and the server is stateless, so a single POST per JSON-RPC message is enough. No session to keep alive, no SSE stream to hold open. An unauthenticated call shows the challenge:
Follow the chain in the diagram above to a Bearer token, then call a tool:
Every result is JSON. Read only tools are annotated as such in the protocol, so a client can distinguish a lookup from a write without guessing, and some tools attach a next_actions array naming the tool to call next. Building on those two signals gives you the guided path the web app offers, without hardcoding a workflow.