Skip to main content
Glama
hernaninverso

eleata-verify-mcp

README.md
# 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](https://eleata.io) 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

```json
{
  "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.

TDQS

A3.7/5.0

Scored across 3 tools

Disambiguation2/5

verify_claim and verify_strict are nearly identical except for an abstention threshold, making it unclear which to choose. check_groundedness overlaps in purpose (checking support in context/evidence). Agents will struggle to differentiate without additional heuristics.

Naming Consistency3/5

Two tools use 'verify_' (one with noun 'claim', one with adjective 'strict'), and one uses 'check_groundedness'. The verbs 'check' and 'verify' are synonymous but break pattern consistency. The mix of noun and adjective styles is acceptable but not uniform.

Tool Count4/5

With 3 tools, the count is slightly low but appropriate for a narrowly scoped verification server. Each tool serves a distinct mode (normal, strict, groundedness), so no tool feels superfluous.

Completeness3/5

Covers the core verification operations but lacks any tool for evidence retrieval, source management, or batch verification. The strict version adds nuance but also indicates incomplete coverage of confidence levels. Agents may need to work around missing functionality.

Maintenance

ActivityInactive
ResponsivenessNo issues