MCP Safety Board
MCP Safety Board · method · rubric v0.1

How the board grades a server

The board grades safety hygiene, not security. A high grade means the server's public code and docs follow the habits that make an MCP server easier to use safely. It is not a security audit, not a penetration test and not a guarantee that a server has no vulnerabilities.

What is measured

Each server is checked once, at the commit shown on its page, using only static reading of the repository. No server code is installed or run.

1. Tool safety labels (annotations)

The MCP spec lets each tool declare hints: readOnlyHint, destructiveHint, idempotentHint and openWorldHint. Clients use them to decide which tool calls need your confirmation. The board counts:

  • Tools: tools the server registers in its source code.
  • Tools labelled: tools that declare at least one of the four hints, or pass an annotations object.
  • Risky tools (destructive-like): tools whose name or description contains one of: delete, remove, drop, write, update, exec/execute, run, eval, kill, revoke.
  • Risky tools labelled: risky tools that declare destructiveHint or readOnlyHint. Either value counts: the tool has told the client what it does.

2. Pinned install instructions

If the README tells users to launch the server with npx, bunx, uvx or pnpm dlx, does the command pin a version (pkg@1.2.3, pkg==1.2.3)? @latest counts as unpinned. "Pinned" means every launch snippet is pinned; "not pinned" means at least one is not; "no snippet" means the README has no such command.

3. Hygiene findings from mcp-audit

Findings from mcp-audit, a static scanner for Python and JS/TS source plus MCP config files of any repo. Only these rules count toward the grade:

RuleWhat it meansSeverity used
MCP002An MCP config committed in the repo launches a package with no version.Medium
MCP003An MCP config committed in the repo uses a moving tag (@latest, :latest).Medium
MCP004A committed config gives a filesystem server a broad root (/, ~, a whole drive).High
MCP005A committed config runs a docker server privileged or with host mounts.High (Critical counted as High)
MCP006A committed config points at a remote MCP server over plain http:// (not localhost).High
MCP105A tool returns raw fetched web content to the model.Info (no points)
MCP106A destructive tool has no annotations.Low (recorded only; already counted in part 1)

Files under test(s)/, example(s)/, sample(s)/, demo(s)/ and fixtures/ are ignored.

Score

Every server starts at 100. Points are taken off as follows:

ItemPenalty
Each risky tool with neither destructiveHint nor readOnlyHint−2 each, at most −30
Fewer than 50% of all tools carry any annotation−10
README launch snippet without a pinned version−10
Hygiene finding, High−8 each
Hygiene finding, Medium−4 each
Hygiene finding, Low (MCP106 excluded, see above)−1 each
Cap on all hygiene findings togetherat most −40; at most 3 findings per rule are counted

The score never goes below 0.

GradeScore
A90 and above
B80 to 89
C70 to 79
D60 to 69
Fbelow 60

Who gets a grade

  • Only repositories that are MCP servers (they register tools) and are not archived or forks.
  • Only servers where at least one tool was counted with confidence (at least one tool, and most tools resolved to a name). Servers that build their tool list at runtime (from an OpenAPI file, a config file, a class registry) are listed as "not graded".
  • Servers written in languages other than Python and JS/TS are graded on parts 1 and 2 and on the config rules of part 3; their source-code rules were not run.

Security issues are handled separately

The scanner also has rules for code-level vulnerabilities (shell, path, SQL, URL, eval, secrets). These do not appear on the board. Anything that looks exploitable is checked by hand and, if real, reported privately to the maintainers first. A server with an open, undisclosed issue is left off the public page until it is fixed or the disclosure window has passed. Its hygiene grade is not changed by it.

If you are a maintainer and want to know whether anything was reported for your server, ask privately via X @AbheeshtTr10795. Please do not post security details in a public issue.

Known limits

  • Tool counting and annotation detection are pattern-based, not a full parser. Tools whose annotations are built by a helper in another file may be counted as unannotated.
  • The risky-word list is a heuristic: run_query counts, reset_branch does not.
  • A pinned README does not prove that the published package is safe; it only means users get the version the maintainer documented.
  • Grades are a snapshot of one commit; maintainers can ask for a re-check at any time.

Re-checks

  • Anyone can ask for a re-check from a server's page ("Request a re-check" opens a GitHub issue). The server is scanned again at its latest commit with the same rubric, and the page shows the new commit and grade.
  • If a count is wrong (for example, annotations built in a helper file), say where in the issue. Confirmed miscounts are corrected by hand.
  • Grades are never changed for any other reason, and nobody can pay for a grade.

Snapshot

  • Snapshot of November 2026: the 317 most-starred open-source MCP servers on GitHub that were checked; 47 graded, 270 not graded.
  • 47 of the 47 grades on this board were also checked by hand.
  • Each server page links to the exact commit that was read. Download the data (JSON).