What is Count?
Count is a collaborative platform for AI-powered analytics and research. Connect it to Claude, ChatGPT, or any assistant that supports MCP, and it splits an AI answer into individual claims, draws the reasoning chain behind each one, and flags hallucinations and assumptions, so you can check the work before you share it.
What is a reasoning chain?
A reasoning chain is a map of how an AI reached a claim. Count draws it as a graph under the claim: the prompt, what the model already believed, each search, query, or calculation, and the source the claim rests on. When a step is wrong or missing, the chain shows where.
Kestrel is winning retail deals since it cut its enterprise price by 20% in March.
- PromptWrite a competitive brief on Kestrel.
- Model knowledgeRecalls Kestrel’s pricing from training data.
- Web searchNo source mentions a price cut.
- HallucinationUnsupported. Source it or cut it.
Count reads the whole conversation, not only the final reply, so the chain includes the searches the assistant ran, the files it read, and the corrections you made along the way. Each claim gets a label:
- Citation. The claim traces to a source you can open: a document, a web page, or a row of data.
- Assumption. The claim rests on something nobody checked. It may be fine; it is worth knowing.
- Model knowledge. The model said it from memory. That knowledge stops at its training date and can be out of date.
- Hallucination. No source supports the claim, or a source contradicts it.
The point is not a score. It is being able to say why an answer is what it is, and to show a colleague the same thing.
How do I check whether ChatGPT’s or Claude’s analysis is right?
Check the claims, not the whole answer. Split the response into the individual statements it depends on, then confirm what each one rests on: a source you can open, a query you can re-run, or an assumption you can test. Count does this for you and links every claim to its evidence.
Checking an answer by hand
- 01Find the load-bearing claims. Numbers, names, dates, and anything that says one thing caused another. If a claim were wrong, would the conclusion change? Those are the ones to check.
- 02Ask where each one came from. A source, a calculation, the model’s memory, or an assumption. Ask the assistant directly, and treat “I recall” or no answer as unsourced.
- 03Open the sources. Check that the page exists and says what the answer claims it says. Misattributed quotes and invented citations look exactly like real ones.
- 04Re-run the numbers. Repeat any calculation, and re-run any query against the real data, not a sample pasted into the chat.
- 05Check the conclusion, not just the facts. A stack of true sentences can still add up to a conclusion nothing supports. Follow the steps between the facts and the recommendation.
Checking an answer with Count
With Count connected to your assistant, ask it to audit the answer. Count splits the answer into claims, traces each one through the conversation to its source, and labels what it finds. You get a link to a canvas: click a claim to see the quote, the query, or the note that nothing was checked. Then decide what to keep, what to fix, and what to send. You can click through an audited answer without signing up.
What is Count’s MCP server?
Count's MCP server connects your AI assistant to Count. Add it once to Claude, ChatGPT, or any app that supports the Model Context Protocol, then ask Count to audit an answer from inside the conversation. Count replies with a link to a canvas holding every claim and its reasoning chain.
The Model Context Protocol (MCP) is the open standard AI assistants use to connect to other tools. Adding Count as an MCP server gives your assistant a way to hand its work to Count, so you audit answers where you already ask questions, not in a new app. Set it up once:
- 01Copy your MCP URL. In Count, it is under Integrations, in MCP settings. The MCP server is part of the Pro and Enterprise plans.
- 02Add the URL to your assistant. Paste it wherever your AI app accepts a remote MCP server. In Claude, that is a custom connector.
- 03Ask Count when an answer matters. Mention Count in the conversation, for example “audit this with Count.” There is no special prompt or format; Count reconstructs the reasoning from the thread.
- 04Open the canvas. Click any claim to see what it stands on, then invite your team to comment on it, correct it, or take the analysis further.
Count also connects outward to MCP servers and databases, so its own agent can work with the same tools and data your team uses. See integrations.
What else does Count do?
The canvas is where the audit lives, and it is a full analytics workspace. Count connects to your data warehouse and business apps, runs SQL and Python, and has its own agent. Teams use it for exploratory analysis, reports, dashboards, and decks, all checked and edited together in real time.
Every query the agent writes is visible and editable, so analysis gets the same treatment as a written answer: you can see how it got there and change any step. Teammates comment on a claim in place, branch off their own question, and present the finished work from the same canvas.
For how that compares with what you might use today, see Count vs chatbots and Count vs BI tools.
FAQs
No. You keep asking questions in Claude, ChatGPT, or whichever assistant you already use. Count is where you check the answer and share it.
Count does have its own agent, powered by models from Anthropic, OpenAI, and Google, for when you want to keep working on the canvas.
Any assistant that can connect to a remote MCP server, including Claude, ChatGPT, Cursor, and VS Code.
No tool can promise that. What Count does is show what every claim rests on, and flag the ones that rest on nothing, so a person can make the call in minutes instead of re-doing the work.
No. Count imports only the reasoning it needs to check the answer’s key claims, including any feedback you gave the assistant. Personal asides are left out.
You can start on the Free plan, with the canvas, Count’s agent, and one database or MCP connection. Connecting your own assistant through Count’s MCP server is part of Pro and Enterprise, along with unlimited canvases, scheduled reports, and shared skills. See count.co/pricing for every plan.
Yes. Nothing leaves your assistant unless you ask Count to audit it, and Count does not train models on your data. Count is SOC 2 compliant, follows GDPR, and can host and process your data in the US or the EU.
For full details, visit our Trust Center at trust.count.co.