Ground your LLM in real data to cut hallucination.
Hallucination is mostly a grounding problem - the model answers from frozen training instead of the current world. Give it a keyless tool layer that returns the live fact plus where and when it came from.
The call
Two keyless calls in one batch - reality_check verdicts a claim against live data, software_version returns the current fact to cite. No key, no signup.
curl -s https://dynamicfeed.ai/v1/batch \
-H 'content-type: application/json' \
-d '{"calls":[
{"tool":"reality_check","args":{"claim":"the latest stable Python is 3.12"}},
{"tool":"software_version","args":{"product":"python"}}
]}'Sample response
{
"mcpServers": {
"dynamic-feed": {
"url": "https://dynamicfeed.ai/mcp"
}
}
}Why live data
Retrieval beats recall for anything that changes. Point your agent at the MCP endpoint (or POST /v1/batch) and it can pull the current value at answer-time instead of reciting a stale one - the software_version call above returns Python 3.14.6, not the 3.12 a frozen model insists on. Use reality_check as a self-check before the model commits to a current fact. Each response carries an Ed25519 signature and a provenance stamp - source, observed-at, freshness. Read the signature honestly: it is proof-of-existence and tamper-evidence - proof that we reported value X from source S at time T, and that the bytes have not been altered since. It is not a claim that X is true, accurate, or safe to act on. You still judge the source; the signature just makes the record verifiable and hard to fake.
Use it for
- Have the agent fetch the live answer instead of guessing from training
- Self-check a draft answer with reality_check before it reaches the user
- Attach a citeable source and timestamp to every grounded fact
- Catch stale software versions, CVE status and service outages at answer-time
FAQ
Does this guarantee my model stops hallucinating?
No. It removes one big cause - answering from stale training - by handing the model a live, citeable fact to ground on. The signature proves we reported that value at that time from that source and that it was not tampered with; it is not a claim the value is true. You still decide whether to trust the source.
Do I need an API key?
No. The MCP endpoint and POST /v1/batch are keyless and free for fair use. Point your client at the URL and start calling the 87 tools.
How does the agent know which tool to call?
Connect over MCP and the agent sees all 87 tools with descriptions and picks at runtime - reality_check to verdict a claim, software_version for releases, check_vulnerability for CVEs, current_weather for conditions, and so on across 19 verticals.