diff --git a/tests/web-model-scope.test.ts b/tests/web-model-scope.test.ts
index 0f819c5..cada37b 100644
--- a/tests/web-model-scope.test.ts
+++ b/tests/web-model-scope.test.ts
@@ -93,6 +93,68 @@ test("non-glob patterns fall back to partial id and name containment", () => {
   ).toBe(true);
 });
 
+test("character classes match one member, like minimatch", () => {
+  expect(
+    modelMatchesScopePattern("zai/glm-5.[23]", "zai", { id: "glm-5.2" }),
+  ).toBe(true);
+  expect(
+    modelMatchesScopePattern("zai/glm-5.[23]", "zai", { id: "glm-5.3" }),
+  ).toBe(true);
+  expect(
+    modelMatchesScopePattern("zai/glm-5.[23]", "zai", { id: "glm-5.4" }),
+  ).toBe(false);
+  expect(
+    modelMatchesScopePattern("zai/glm-5.[2-3]", "zai", { id: "glm-5.3" }),
+  ).toBe(true);
+  expect(
+    modelMatchesScopePattern("zai/glm-5.[2-3]", "zai", { id: "glm-5.4" }),
+  ).toBe(false);
+});
+
+test("negated character classes exclude members and the path separator", () => {
+  expect(
+    modelMatchesScopePattern("zai/glm-5.[!23]", "zai", { id: "glm-5.4" }),
+  ).toBe(true);
+  expect(
+    modelMatchesScopePattern("zai/glm-5.[!23]", "zai", { id: "glm-5.2" }),
+  ).toBe(false);
+  expect(
+    modelMatchesScopePattern("zai/glm-5.[^23]", "zai", { id: "glm-5.x" }),
+  ).toBe(true);
+  expect(
+    modelMatchesScopePattern("zai/glm-5.[^23]", "zai", { id: "glm-5.2" }),
+  ).toBe(false);
+  // A negated class never matches "/", so it cannot span the provider boundary.
+  expect(modelMatchesScopePattern("zai[!z]*", "zai", { id: "glm-5.4" })).toBe(
+    false,
+  );
+  // A class whose only member is "/" can never match anything.
+  expect(modelMatchesScopePattern("zai[/]*", "zai", { id: "glm-5.4" })).toBe(
+    false,
+  );
+});
+
+test("unterminated character classes fall back to a literal bracket", () => {
+  expect(
+    modelMatchesScopePattern("zai/glm-5.[2", "zai", { id: "glm-5.2" }),
+  ).toBe(false);
+  expect(
+    modelMatchesScopePattern("zai/glm-5.[2", "zai", { id: "glm-5.[2" }),
+  ).toBe(true);
+});
+
+test("a leading bracket inside a class is a literal member", () => {
+  expect(
+    modelMatchesScopePattern("zai/glm-5.[]23]", "zai", { id: "glm-5.]" }),
+  ).toBe(true);
+  expect(
+    modelMatchesScopePattern("zai/glm-5.[]23]", "zai", { id: "glm-5.2" }),
+  ).toBe(true);
+  expect(
+    modelMatchesScopePattern("zai/glm-5.[]23]", "zai", { id: "glm-5.4" }),
+  ).toBe(false);
+});
+
 test("readEnabledModelPatterns reads enabledModels from a settings file", async () => {
   directory = await mkdtemp(join(tmpdir(), "pi-model-scope-"));
   const settingsPath = join(directory, "settings.json");
diff --git a/web/server/modelScope.ts b/web/server/modelScope.ts
index 03f4c90..e78e0da 100644
--- a/web/server/modelScope.ts
+++ b/web/server/modelScope.ts
@@ -47,7 +47,9 @@ export async function readEnabledModelPatterns(
 /**
  * Convert a glob to a RegExp with minimatch's default semantics for the
  * characters we support: `*` and `?` never cross `/`, and `[...]` character
- * classes pass through. Everything else is literal.
+ * classes (sets, ranges, `!`/`^` negation, a leading `]` as a literal member,
+ * `\`-escaped members, and unterminated `[` treated as a literal) translate
+ * directly. Everything else is literal.
  */
 function globToRegExp(pattern: string): RegExp {
   let source = "";
@@ -56,11 +58,74 @@ function globToRegExp(pattern: string): RegExp {
     if (char === "*") source += "[^/]*";
     else if (char === "?") source += "[^/]";
     else if (char === "\\") source += "\\\\";
-    else source += char.replace(/[.+^${}()|[\]]/g, "\\$&");
+    else if (char === "[") {
+      const charClass = parseCharClass(pattern, index);
+      if (charClass) {
+        source += charClass.source;
+        index = charClass.nextIndex - 1;
+      } else {
+        // Unterminated class: minimatch treats the `[` as a literal.
+        source += "\\[";
+      }
+    } else source += char.replace(/[.+^${}()|[\]]/g, "\\$&");
   }
   return new RegExp(`^${source}$`, "i");
 }
 
+/** Escape a character so it is a literal member of a RegExp character class. */
+function escapeClassMember(char: string): string {
+  return char === "\\" || char === "]" || char === "^" ? `\\${char}` : char;
+}
+
+/**
+ * Translate the `[...]` character class starting at `start` into a RegExp
+ * class source with minimatch's semantics, returning the source and the index
+ * just past the closing `]`. `undefined` when the class is unterminated.
+ */
+function parseCharClass(
+  pattern: string,
+  start: number,
+): { source: string; nextIndex: number } | undefined {
+  let index = start + 1;
+  let negate = false;
+  if (pattern[index] === "!" || pattern[index] === "^") {
+    negate = true;
+    index += 1;
+  }
+  const members: string[] = [];
+  // A `]` immediately after `[` (or the negation) is a literal member.
+  if (pattern[index] === "]") {
+    members.push("\\]");
+    index += 1;
+  }
+  let closed = false;
+  while (index < pattern.length) {
+    const char = pattern[index];
+    if (char === "\\") {
+      const next = pattern[index + 1];
+      if (next === undefined) return undefined;
+      members.push(escapeClassMember(next));
+      index += 2;
+      continue;
+    }
+    if (char === "]") {
+      closed = true;
+      index += 1;
+      break;
+    }
+    members.push(escapeClassMember(char));
+    index += 1;
+  }
+  if (!closed) return undefined;
+  // Minimatch classes never match the path separator: a `/` member can never
+  // match (dropped), and a negated class excludes `/` alongside its members.
+  const literalMembers = members.filter((member) => member !== "/");
+  return {
+    source: `[${negate ? "^/" : ""}${literalMembers.join("")}]`,
+    nextIndex: index,
+  };
+}
+
 function hasGlobCharacter(pattern: string): boolean {
   return /[*?[]/.test(pattern);
 }