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 establishes integrity and attribution under the published key, so altered bytes fail verification. It does not independently prove time, truth, accuracy, or safety. You still judge the source; the signature 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

Related live feeds