An evidence-weighing Model Context Protocol (MCP) server for Claude, ChatGPT and Cursor
An AI assistant writes plausible text. It does not keep a ledger of where each statement came from or how much that source deserves to be trusted. Arbiter is that ledger, and the Model Context Protocol (MCP) is how the assistant reaches it.
What the connection adds
With Arbiter connected, an assistant can do more than summarise your sources. It can:
- Record each source as a named entry with a type, such as regulator, researcher, journalist or anonymous, and a reliability that follows from the type.
- Tie each claim to the source that made it, and each piece of evidence to the claim it supports or opposes.
- Ask Arbiter to judge the investigation. Arbiter weighs the sources by reliability and recency, counts sources marked as repeating one another only once, and returns a verdict per question with a confidence score, plus a status for each claim. Per-claim scores need a paid plan.
- On a paid plan, read back the audit trail: every claim with who made it and how strongly the evidence supports and opposes it.
Each investigation is saved in your account, so you can open it later in the Arbiter console and walk through the reasoning.
Connection details
| Server address | https://api.arbiter.housecarl.cloud/mcp/v1 |
| Sign-in | A sign-in window (Claude and other clients that support OAuth), or an API key in the x-api-key header |
| Transport | Streamable HTTP |
| Assistants | Claude (web, mobile, Desktop, Claude Code), ChatGPT, Cursor, Windsurf, Google Antigravity, and any client that supports remote MCP servers |
Connect now: Use it from Claude takes five minutes and needs no code; free, no card required. Step-by-step setup for every other assistant is in the Quick Start.
What it will not do
Connecting Arbiter does not change the model. The assistant can still misread a source when it extracts the claims. What changes is what the verdict rests on: named sources with visible weights, which you can check. We measure the sources you give us against what they say; if the sources are wrong, the verdict is wrong with them.
Next
Current plans are on the pricing page. If you are building your own software rather than using an assistant, read MCP or REST API: which to use.
Related: Arbiter for researchers · Five tests for an AI fact-checking tool