I use Claude Code for many non-coding-based tasks, and I love most of what it does. Be it subagents, MCPs, or the many commands to streamline your results, Claude Code gets almost everything right! However, there are times when it presents questionable statements as facts — with a worrying level of confidence!
These subtly placed hallucinations can be a costly mistake when you use Claude Code for research, though. Manually fact-checking every major claim in a paragraph isn’t a viable solution, either. Fortunately, I came across a free fact-checking tool called Lenz, and it has changed the way I use Claude Code for research.
Lenz checks the claim
It doesn’t care how sure Claude Code sounded
Lenz is an independent fact-checking tool that you can integrate with coding agents and AI tools like Claude Code and Claude. It serves as an intermediary between the AI model and the content you ultimately use, independently fact-checking the claims it generates. As you can guess, this proposition becomes important when you cannot afford hallucinations or half-baked statements for your research!
We shouldn’t mistake this tool for a search tool like Firecrawl, though. Lenz is a fully fledged fact-checking platform that uses multiple models to assess the veracity of a claim. Instead of relying on what Claude or Claude Code already provides, Lenz can use its independent, multi-model pipeline to provide a score and an indicator. Depending on your requirements, you can also choose between four primitives: extract, assess, verify, and ask.
The extract function can extract the claim from a statement, even when Claude Code confidently asserts it. Instead, it extracts only the claim, not the emotion, and this claim is fact-checked using the multi-model pipeline. Then, depending on the level of fact-checking you need, you can choose between verify (which takes up 90 seconds but provides citations) and assess (which takes 10 seconds but lacks citations).
Despite how complex the mechanism is, setting up Lenz was easy, though.
- OS
-
Web
- Individual pricing
-
Free, Subscription available
Lenz is an independent fact-checker tool you can integrate with AI tools to verify the claims generated by a model
Getting Lenz running inside Claude Code took four lines in a terminal
No config file surgery required
Setting up Lenz inside Claude Code was easier than I expected. I went with the MCP server option, which still requires you to enter the API key in the Terminal. I ran the following command on Claude Code, and the setup was done in a few seconds.
$ claude mcp add –transport http lenz
https://lenz.io/mcp
–header “Authorization: Bearer ${LENZ_API_KEY}”
After this simple setup process, I could use a suitable command to invoke Lenz. I’ve also managed to integrate some of these Lenz skills into subagents and skills I’ve made using Claude Code. For instance, I’ve created a fact-checking subagent that uses Lenz’s extract-and-verify components to reach a consensus on a claim’s veracity.
This subagent has been a great boost to my workflow! Instead of manually running a fact-check whenever required, I can invoke the subagent when Claude Code has produced something. This way, I can ensure that I’m not publishing non-factual just because Claude said it confidently.
I’ve also connected Lenz to Claude AI using the official connector, and it’s been equally great! I can invoke this connector’s fact-checking prowess by mentioning the term “Lenz” in the prompt. As I said earlier, the most significant difference has been on the research front.
I ran the same research task with and without Lenz connected
One version sounded sure. Only one of them was right
To be completely clear, I don’t use Lenz all the time when Claude Code is running. Doing so would be a waste of tokens and credits. Instead, I bring the fact-checking subagent to the equation only when the situation absolutely demands it. Here’s an instance:
Let’s say that I’m researching something where numbers have to be absolutely correct. In this case, I wouldn’t be comfortable with Claude providing the data because it could state untrue or partially true statements as facts. On the contrary, once I bring Lenz into the equation, I can ask Lenz to extract all claims that contain a number. This way, I get to ensure that the numbers are as accurate as possible.
It’s nice that Lenz also offers options for follow-up questions about the fact-checking instance. These responses are also from the multi-model pipeline, and you don’t have to worry about accuracy. The verify command takes things further by running a multi-model debate on the topic, but the process takes time.
Lenz offers a free plan that gives you 100 credits every month, and they’ve been sufficient for the workflow I mentioned. If you want to check hundreds or thousands of claims every week, you might need a paid plan from Lenz.
Lenz doesn’t fix the confidence problem. It works around it
While Lenz makes my research output more reliable, it’s not a perfect solution. Its fact-checking pipeline performs better than Claude Code’s, but it cannot access paywalled content or other restricted sources. In those cases, I’ve come across an inconclusive message. Yet, it is an exception, not the norm. For most people, adding this free MCP to Claude Code — or other agentic AI tools — is an effective way to boost research reliability within Claude.