mcp-evalSAFE
🧐 Privacy-preserving MCP server evaluation: deterministic CI probe battery (18 probe kinds, readiness score, SARIF) plus production friction capture with findings that verify and close
Overview
From the repository's own README, as read at the audited commit. Badges and raw HTML are left out.
[](LICENSE)
mcp-eval is an agnostic evaluation tool for MCP servers, with two complementary tracks:
- Deterministic benchmark battery — run a strict, manifest-driven probe
suite against any MCP server (stdio or Streamable HTTP) and get a pass/fail verdict with fixed failure reasons. No LLM in the loop, no network services, fully reproducible: same server, same manifest, same result. Built for CI gating while you develop a server.
- Friction capture — run your server behind a privacy-preserving shim in
production-like agent sessions, index what it actually costs agents, and promote repeated friction into queryable findings a development agent can act on. Findings close only after three consecutive green verification runs and reopen automatically on regression.
Capture journals and probe summaries omit raw payloads, error prose, and credentials. Store records under /store/ are eligible for sharing after reviewing annotation notes and the remaining files. Manifests, corpus target files, and retained corpus evidence are private artifacts outside that boundary; see Privacy and authorization.
Documentation: cavi-ai.xyz/docs/mcp-eval.
Install
brew install cavi-ai/tap/mcpeval npm install --global @cavi-ai/mcp-eval cargo install mcpeval --locked
Each installs mcpeval and the demo server mcpeval-demo. Homebrew (macOS and Linux, x64/arm64) and npm (also Windows x64) install the prebuilt release archive after checking its pinned SHA-256; the archives and their checksums are also on the releases page. Building from a checkout works too:
cargo build --release
The binaries are target/release/mcpeval and target/release/mcpeval-demo (.exe on Windows).
Quick
fce0d5110a8eOBSERVED · 2026-10-11Connect
Built from this server's own package name, version and transport as found in its source — not copied from anyone's documentation, so it cannot drift against a page we do not control.
claude mcp add mcp-eval -- npx -y @cavi-ai/[email protected]
{
"mcpServers": {
"mcp-eval": {
"command": "npx",
"args": [
"-y",
"@cavi-ai/[email protected]"
]
}
}
}Exposed tools (1)
1 read · 0 write · 0 destructive.
| Tool | Risk | Description |
|---|---|---|
read_state | read | Read the fixture dataset |
Trust audit
SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.
| Layer | What it checks | Result |
|---|---|---|
| L0 | Provenance & inventory | PASS |
| L1 | Static analysis of the code | PASS |
| L2 | Instruction surface (what it tells the agent) | PASS |
| L3 | Class-specific surface | PASS |
| L4 | Behavioural (sandbox) | SKIPPED |
What the source does
- Filesystem
- declared (1 observation(s))
- Network
- declared (6 observation(s))
- Shell
- declared (1 observation(s))
- Dependencies
- pinned
- Secrets in source
- none-found
Findings (12)
.pr-gates
import { checkEvidence } from "../../scripts/corpus/check-evidence.mjs";const corpusFixture = JSON.parse(await readFile(new URL("../../tests/fixtures/corpus-v3.json", import.meta.url), "utf8"));const metadataCases = JSON.parse(await readFile(new URL("../../tests/fixtures/corpus-v3-metadata-cases.json", import.meta.url), "utf8"));for (const change of [{ root: "relative" }, { sha256: "unknown" }, { executable: "../outside" }, { executable: "/outside" }, { entrypoint: "a/../../outside" }, { executable: "runtime\\outside" }]) {new URL("../../.github/workflows/publish.yml", import.meta.url),--probe discovery-cost --url http://127.0.0.1:8080/mcp
--listen 127.0.0.1:8090 --upstream http://127.0.0.1:8080/mcp
url: http://127.0.0.1:8080/mcp
--url http://127.0.0.1:8080/mcp
--upstream http://127.0.0.1:8080/mcp
`--format markdown` renders a pull-request-ready report: the readiness score with its standard and a static shields.io badge URL encoding only the score, an area table, the surface, *Lost points* with
Gates applied: no_behavioural_pass.
fce0d5110a8efull audit observations/trust-audit/mcp-server/cavi-ai__mcp-eval.json · Report an issue / request a re-scanAudit history
Every audit this server has had. A grade with a past is a grade somebody is still checking.
| Date | Source | Verdict | Grade | Score | Change |
|---|---|---|---|---|---|
| 2026-10-11 | fce0d5110a8e | SAFE | B | 89 | first audit |
Questions
What is the mcp-eval MCP server?
🧐 Privacy-preserving MCP server evaluation: deterministic CI probe battery (18 probe kinds, readiness score, SARIF) plus production friction capture with findings that verify and close
What tools does mcp-eval expose?
1 in total: 1 read-only, 0 that write, and 0 that can delete or overwrite. Every one is listed on this page with its risk.
Is mcp-eval safe to connect to an agent?
The audit found nothing in the source that contradicts what it says it does, and graded it B (89/100). Grade A is held back for packages that have also passed a sandboxed behavioural run, which is why a clean server reads B.
What credentials does mcp-eval need?
No credential environment variables were found in its source, so it appears to need none.
How does mcp-eval run?
It speaks stdio and streamable-http, so it runs as a local process your client starts. It is published on npm as @cavi-ai/mcp-eval at 0.6.0.
How current is this page?
The grade is for one exact copy of the source (fce0d5110a8e), read on 2026-10-11. The repository is watched and re-audited when it changes.