Results / model-finder-mcp
model-finder-mcp
Sealmark's automated test of model-finder-mcp 1.3.7 on 2026-11-20 gave it grade B, its lowest test-area grade (protocol B and effects A). Only 2 of 10 tools could be called with default settings, so the deeper test areas cover only those tools. Of its 6 checks, 4 passed and 2 partial. 3 more checks are advisory, and advisory checks never change a grade: 1 failed and 2 not applicable. The run used Sealmark 1.0.1 with the standard settings.
Grade B
The server's grade is its lowest test-area grade. How grading works
Only 2 of 10 tools could be called with default settings, so the deeper test areas cover only those tools.
Grade by test area
Protocol checks
| Check | Verdict | What happened |
|---|---|---|
| protocol Valid MCP handshake and tool schemas | pass | handshake ok, 10 well-formed tool(s) Evidence for protocol: 1 line
|
| tool-poisoningweak signal No tool-poisoning indicators in tool metadata | partial | 1 weak indicator(s) worth reading, none conclusive Weak signal, not a security finding. Some wording in the tool text is worth a read, but none of it is conclusive. A partial counts as half a point. |
| robust-input Survives malformed tool input | partial | 0 internals leak(s), 1 invalid call(s) accepted Evidence for robust-input: 1 line
|
| secret-echocritical No environment secret echoed in responses | pass | canary secrets absent from tools/list and 52 captured response(s), 1 of them successful result(s) |
| instructions-hygiene (advisory) Server instructions name only tools that exist | fail | the server instructions names 1 tool(s) the server does not offer (eval_models) |
| refusal-quality (advisory) Refusals say what to do next | n/a | no refusal was recorded in this run, so there is nothing to judge |
Effects checks
| Check | Verdict | What happened |
|---|---|---|
| annotation-honestycritical Read-only tools change nothing | pass | modelfinder_audit_models declared read-only and changed nothing in the observed directories during a benign call (modelfinder_draft_eval_examples not judged: the control call failed) |
| openworld-honesty Closed-world tools stay on the machine | pass | 2 closed-world tool(s) called; nothing connected to the loopback listener (path and git tools cannot be pointed at it, so this confirms only that nothing connected out during the call) |
| undeclared-writes (advisory) Unannotated tools leave the sandbox unchanged | n/a | every tool carries annotations, nothing to check |
Each check id links to what it tests and how to fix a failure. The same text is in sealmark explain <id>.
Under a check, Evidence shows what it probed (sent or inspected) and what it observed in this run: up to 6 lines per check, each cut at 240 characters, for checks that are not security checks. The text is shortened, and paths and anything that looks like a secret or token are replaced. Run sealmark run --format json yourself for every line in full.
By tool
Which checks called each tool with a valid call in this run, and which did not exercise it. A check can leave a tool out for several reasons: a limit Sealmark sets, a tool left out by design, input the run could not fill in, or a call that returned an error. The CLI report gives the reason. A verdict belongs to the whole check, not to one tool: a failed check that called three tools does not say which one failed. Security checks are left out of this table.
| Tool | Called by | Not exercised by |
|---|---|---|
| modelfinder_audit_models | No check listed here called it with a valid call. | None |
| modelfinder_draft_eval_examples | No check listed here called it with a valid call. | None |
| modelfinder_create_eval_config | No check listed here called it with a valid call. | None |
| modelfinder_validate_eval_config | No check listed here called it with a valid call. | None |
| modelfinder_check_output | No check listed here called it with a valid call. | None |
| modelfinder_rank_eval_results | No check listed here called it with a valid call. | None |
| modelfinder_generate_routing | No check listed here called it with a valid call. | None |
| modelfinder_estimate_cost | No check listed here called it with a valid call. | None |
| modelfinder_estimate_memory | No check listed here called it with a valid call. | None |
| modelfinder_scan_sensitive_data | No check listed here called it with a valid call. | None |
What was not covered
- Database: Not applicable. The start command takes no database connection, so the database tests did not run. ({dsn}, {host} and the like are placeholders that Sealmark fills in with its own test database.)
- HTTP: Not applicable. The server is started over stdio; the http test area needs server.url
- Files: Not applicable. The server config has no {sandbox} placeholder, so the harness cannot tell which directory the server is allowed to use (a server with a wider root would read the canary legitimately); pass {sandbox} as its directory argument, or, if the server has no root argument and works in its working directory, set server.rootless: true
- Git: Not applicable. No tool is classified as taking a git repository path with medium confidence or higher
- Network: Not applicable. No tool is classified as taking a URL with medium confidence or higher
- SQLite: Not applicable. No tool takes a SQL string
- Tools that a server lists only for clients that declare roots, sampling or elicitation capabilities are not seen: the harness declares none
- Tool "modelfinder_create_eval_config" is covered by the protocol checks only: free-text parameter "task": no path, URL, command, SQL or repository signal; its required parameter "examples" cannot be filled by the harness; write a probes: entry to call it
- Tool "modelfinder_validate_eval_config" is covered by the protocol checks only: no free-text parameter: every string parameter is an enum; its required parameter "config" cannot be filled by the harness; write a probes: entry to call it
- Tool "modelfinder_check_output" is covered by the protocol checks only: free-text parameter "text": no path, URL, command, SQL or repository signal; add it under tools: to include it
- Tool "modelfinder_rank_eval_results" is covered by the protocol checks only: no free-text parameter: every string parameter is an enum; its required parameter "results" cannot be filled by the harness; write a probes: entry to call it
- Tool "modelfinder_generate_routing" is covered by the protocol checks only: string-array parameters only: none names a path, URL, command, SQL or repository; its required parameter "results" cannot be filled by the harness; write a probes: entry to call it
- Tool "modelfinder_estimate_cost" is covered by the protocol checks only: no free-text parameter: every string parameter is an enum; add it under tools: to include it
- Tool "modelfinder_estimate_memory" is covered by the protocol checks only: no free-text parameter: every string parameter is an enum; add it under tools: to include it
- Tool "modelfinder_scan_sensitive_data" is covered by the protocol checks only: free-text parameter "text": no path, URL, command, SQL or repository signal; add it under tools: to include it
- The sandbox is a convention the harness follows, not a security boundary. A malicious server runs with your user's privileges, sees the real file system and network, and can ignore cwd, HOME and PATH. Run untrusted servers in a throwaway VM or container.
- Prompt injection carried in tool output (only tool metadata is scanned)
- Authentication and transport security
- Audit logging (not measured by these test areas; only the db test area has an advisory audit check)
Grade over time
Since the run on
Same as that run: the same grade in every test area, the same server version (1.3.7), the same Sealmark version (1.0.1).
- Bv1.3.7Sealmark 1.0.0
- Bv1.3.7Sealmark 1.0.1
- Bv1.3.7Sealmark 1.0.1
Badge for the README
The badge shows this automated result for version 1.3.7. It is a link to this page, not a certificate, and it changes when the next re-test does.
[](https://sealmark.dev/results/model-finder-mcp/)
Files: badge SVG, card SVG, shields endpoint JSON.
Package facts
- License
- MIT
- Latest release
- 1.3.7, published 3 Nov 2026 (the version graded on this page)
- Repository
- Abhiindigloo/EvalHarness
- Last push
- 21 Nov 2026
- Open issues and pull requests
- 5
From PyPI and GitHub, checked 22 Nov 2026. Plain facts about the package; they play no part in the grade.
Add it to an MCP client
This starts version 1.3.7, the version graded on this page, with the start command from the test run. For a client that reads an mcpServers JSON file:
{
"mcpServers": {
"model-finder-mcp": {
"command": "uvx",
"args": [
"model-finder-mcp==1.3.7"
]
}
}
}Or add it to Claude Code:
claude mcp add model-finder-mcp -- uvx model-finder-mcp==1.3.7
The test run started the server with no API keys or other secrets, and with a cleaned environment. If the server needs some, its own documentation says how to pass them.
Run it yourself
This repeats the run with the same Sealmark version and the standard settings, starting the server the same way:
npx -y sealmark@1.0.1 run --start "uvx model-finder-mcp==1.3.7"
To confirm your run used the standard settings, add --format json --out report.json to the command. In the report, config.standard should be true and config.hash should start with sha256:d13e7893a0ac, as they did in this run. If either differs, so did the settings, and the verdicts may too (STANDARD and CUSTOM runs).
To keep the setup, put the start command in sealmark.yaml as the server.
Grade alerts
One email when the grade of model-finder-mcp changes after a re-test. Nothing else, and every email has a link to stop them.
Or follow it in a feed reader: the Atom feed lists each move of the overall grade that came with a different version of this server, tested with the same Sealmark version as the graded run before.
Each of 613 verdicts was checked independently, and Sealmark was wrong on some: how reliable Sealmark is.
The issue we opened No reply yet
Sealmark opened a public issue on this project's repository on 17 Nov 2026 to share this grade. No reply yet. A reply is optional.
Read the issue on GitHub. State as of the last re-test run.
Is this wrong?
Reply on the GitHub issue or email disclosure@sealmark.dev and we will recheck it. Maintainers can ask for their server to be removed.
If we were wrong, we correct the page and add your reply. Security findings go to the maintainer privately first; see the disclosure policy.