Eleata Verify Mcp
About
Grounding / hallucination guard: verify a claim against evidence
Details
- Author
- hernaninverso
- Downloads
- 220
- Categories
- Other
Jump to
- Natural-language inference verdict (Supports / Refutes / Not Enough Evidence)
- Abstained flag for low-confidence predictions
- Three tools: verify_claim, check_groundedness, verify_strict
- Multilingual model (mDeBERTa-v3-mnli-xnli) behind the API
- Two API channels: direct (Bearer token) and RapidAPI marketplace
- Confidence uncalibrated on public channels; rely on verdict and abstained flag
Setting up with Highlight
This MCP is not yet compatible with Highlight’s one-click setup. However, you can still use it with Highlight by following these steps:
- Download and install Highlight from highlightai.com/download
- Navigate to the plugins tab and select "Add Custom Plugin"
-
Configure the plugin with the settings below
Plugin Name
Eleata Verify McpCommand (node, npx, python, etc.)Please refer to the README for specific instructions on how to obtain API keys or other required environment variables.
- Enable "Start Automatically" if you want the plugin to start when Highlight launches
From the repository
Install via npx -y eleata-verify-mcp and set the EVERIFY_API_KEY environment variable (get a key at eleata.io). Configure in your MCP client as an mcpServers entry. Invoke one of three tools: verify_claim(claim, evidence), check_groundedness(answer, context), or verify_strict(claim, evidence) for a higher abstention threshold.
verify_claim
Fact-check a claim against a piece of evidence using natural-language inference. Returns a grounded verdict — Supported, Refuted, or Not Enough Evidence — with a confidence (0..1) and an `abstained` flag. Use it to check whether a statement is actually backed by a source before trusting or repeating it. If `abstained` is true, treat the result as 'cannot verify — do not rely'.
check_groundedness
Hallucination guard for RAG / agent answers. Pass the model's generated answer and the retrieved context; returns whether the answer is grounded in the context. Treat any verdict other than 'Supported', or `abstained=true`, as 'ungrounded / likely hallucination' and have the agent retract or re-retrieve.
verify_strict
Same as verify_claim but with a raised abstention threshold — abstains more readily. Use when a wrong 'Supported' is costly (compliance, legal, medical-adjacent) and the agent must not over-assert.
Claude Desktop / Cursor
Paste into your MCP client config file to install this server.
{
"mcpServers": {
"eleata verify mcp": {
"eleata-verify": {
"command": "npx",
"args": [
"-y",
"eleata-verify-mcp"
],
"env": {
"EVERIFY_API_KEY": "your_key"
}
}
}
}
}McpServers
{
"eleata-verify": {
"command": "npx",
"args": [
"-y",
"eleata-verify-mcp"
],
"env": {
"EVERIFY_API_KEY": "your_key"
}
}
}eleata Claim Verifier — MCP server
A grounding / hallucination guardrail for AI agents. Give it a claim and supporting
evidence; it returns whether the evidence Supports / Refutes / gives Not Enough Evidence
for the claim (natural-language inference), with a confidence and an abstained flag for
low-confidence cases. Wraps the hosted eleata Claim Verifier.
> Use it as a fact-check / RAG hallucination guard: before an agent trusts or repeats a
> statement, check it against its source. Treat abstained=true (or any verdict other than
> Supported) as "do not rely".
Tools
| Tool | What it does |
|------|--------------|
| verify_claim(claim, evidence) | Is the claim supported by the evidence? Verdict + confidence + abstained. |
| check_groundedness(answer, context) | RAG guard: is the model's answer grounded in the retrieved context? |
| verify_strict(claim, evidence) | Same, with a higher abstention threshold — for compliance/legal where a wrong "Supported" is costly. |
Setup
{
"mcpServers": {
"eleata-verify": {
"command": "npx",
"args": ["-y", "eleata-verify-mcp"],
"env": { "EVERIFY_API_KEY": "your_key" }
}
}
}
Get a key at <https://eleata.io/checkout?p=verifypro> (paid channel, sent as Authorization: Bearer).
For the RapidAPI marketplace channel, set EVERIFY_RAPIDAPI=1 and EVERIFY_API_BASE to the RapidAPI host
(the key is then sent as X-RapidAPI-Key).
Notes
Confidence is uncalibrated on the public channels (not a probability) — rely on the verdict and the
abstained flag, not the raw number. The engine is a multilingual NLI model (mDeBERTa-v3-mnli-xnli).
verify sends your claim + evidence to the hosted API; see <https://eleata.io/privacy/>.
MIT licensed.
Sign in to leave a review
Use Google, GitHub, or an email account so ratings stay tied to real people.
No reviews posted yet.
