diff --git a/AGENTS.md b/AGENTS.md
index f76e2ecb7..0760c585f 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -33,6 +33,26 @@ the Ruby VM to the Rust crate logic.
 - `lib`: The rest of the Ruby code
 - `test`: Ruby test files
 
+### MCP launcher
+
+`exe/rubydex_mcp` is a thin Ruby wrapper that boots a Bundler context before `exec`ing the precompiled
+`rubydex_mcp` Rust binary. Its only job is to discover index roots that require Ruby/Bundler context — the
+Rust server still performs the actual indexing pass.
+
+`lib/rubydex/mcp_server_bridge.rb` (`Rubydex::MCPServerBridge`) holds the launcher logic:
+
+- It is launcher-only. **Do not require it from `lib/rubydex.rb`** — the main Ruby API owns graph construction
+  through `Rubydex::Graph`.
+- `setup_bundler_context` activates the workspace's Bundler environment. It mutates `ENV["BUNDLE_GEMFILE"]`
+  (only if unset), which is acceptable in the launcher since it execs immediately, but long-lived callers
+  (e.g. tests) must save and restore that variable.
+- `compute_index_paths` returns the workspace path(s) plus every Bundler dependency `require_path`, so the
+  Rust server indexes both the project and its gem dependencies. Only the first argv entry drives the Bundler
+  context, but all argv entries become index roots.
+- Index roots may overlap (a dependency nested under the workspace, or the workspace's own gemspec resolving
+  back to the root). This is harmless: the graph keys documents by `UriId` and overwrites on re-index, so a
+  doubly-discovered file costs only wasted indexing work, never duplicate declarations.
+
 ### Naming Conventions
 
 The C extension uses prefixed function names to distinguish between abstraction layers:
diff --git a/exe/rubydex_mcp b/exe/rubydex_mcp
index bca33de47..cf928ec32 100755
--- a/exe/rubydex_mcp
+++ b/exe/rubydex_mcp
@@ -1,11 +1,11 @@
 #!/usr/bin/env ruby
 # frozen_string_literal: true
 
-require "rbconfig"
+$LOAD_PATH.unshift(File.expand_path("../lib", __dir__))
 
-host_os = RbConfig::CONFIG.fetch("host_os")
-executable = host_os.match?(/mswin|mingw|cygwin/) ? "rubydex_mcp.exe" : "rubydex_mcp"
-binary = File.expand_path("../lib/rubydex/bin/#{executable}", __dir__)
+require "rubydex/mcp_server_bridge"
+
+binary = Rubydex::MCPServerBridge.binary_path
 
 unless File.executable?(binary)
   abort(<<~MESSAGE.chomp)
@@ -14,4 +14,13 @@ unless File.executable?(binary)
   MESSAGE
 end
 
-exec(binary, *ARGV)
+args = if Rubydex::MCPServerBridge.passthrough?(ARGV)
+  ARGV
+elsif Rubydex::MCPServerBridge.setup_bundler_context(ARGV.first || Dir.pwd)
+  # The wrapper only discovers paths that require Ruby/Bundler context. Rust still performs the indexing pass.
+  Rubydex::MCPServerBridge.compute_index_paths(ARGV)
+else
+  ARGV.empty? ? [Dir.pwd] : ARGV
+end
+
+exec(binary, *args)
diff --git a/lib/rubydex/mcp_server_bridge.rb b/lib/rubydex/mcp_server_bridge.rb
new file mode 100644
index 000000000..aec4b889a
--- /dev/null
+++ b/lib/rubydex/mcp_server_bridge.rb
@@ -0,0 +1,80 @@
+# frozen_string_literal: true
+
+require "rbconfig"
+
+module Rubydex
+  # Launcher-only bridge for `exe/rubydex_mcp`.
+  #
+  # Do not require this file from `lib/rubydex.rb`. The main Ruby API owns graph construction through
+  # `Rubydex::Graph`; this file exists only so the executable wrapper can inspect Bundler before it execs the Rust MCP
+  # server. The Rust server still performs the indexing pass.
+  module MCPServerBridge
+    PASSTHROUGH_ARGUMENTS = ["-h", "--help", "-V", "--version"].freeze
+
+    extend self
+
+    def executable_name
+      host_os = RbConfig::CONFIG.fetch("host_os")
+      host_os.match?(/mswin|mingw|cygwin/) ? "rubydex_mcp.exe" : "rubydex_mcp"
+    end
+
+    def binary_path
+      File.expand_path("bin/#{executable_name}", __dir__)
+    end
+
+    def passthrough?(argv)
+      argv.any? { |arg| PASSTHROUGH_ARGUMENTS.include?(arg) }
+    end
+
+    # Activates the workspace's Bundler environment so dependency gems can be discovered.
+    #
+    # Side effect: sets ENV["BUNDLE_GEMFILE"] (only if unset) to the workspace Gemfile, since
+    # `bundler/setup` resolves against that variable. This permanently mutates process env — acceptable
+    # in the launcher, which execs the Rust binary immediately afterwards, but callers in long-lived
+    # processes (e.g. tests) should save and restore ENV["BUNDLE_GEMFILE"] themselves.
+    def setup_bundler_context(workspace_path)
+      root = File.expand_path(workspace_path)
+      root = File.dirname(root) if File.file?(root)
+
+      gemfile = File.join(root, "Gemfile")
+      ENV["BUNDLE_GEMFILE"] ||= gemfile if File.file?(gemfile)
+
+      require "bundler/setup"
+      require "bundler"
+
+      true
+    rescue LoadError, Bundler::BundlerError => e
+      warn("Warning: failed to load Bundler context: #{e.message}")
+      false
+    end
+
+    # Builds the full set of index roots: the workspace path(s) from argv plus every Bundler
+    # dependency require_path.
+    #
+    # Contract note: only the first argv entry drives the Bundler context (see the call to
+    # `setup_bundler_context(ARGV.first || Dir.pwd)` in `exe/rubydex_mcp`), but ALL argv entries
+    # become index roots here. So dependency paths are resolved relative to the first path's
+    # Gemfile, while every requested path is still indexed.
+    def compute_index_paths(argv)
+      paths = argv.empty? ? [Dir.pwd] : argv
+      paths = paths.map { |path| File.expand_path(path) }
+
+      paths.concat(dependency_index_paths)
+      paths.uniq
+    end
+
+    def dependency_index_paths
+      Bundler.load.specs.flat_map do |spec|
+        spec.require_paths.filter_map do |path|
+          next if File.absolute_path?(path)
+
+          full_path = File.join(spec.full_gem_path, path)
+          full_path if File.directory?(full_path)
+        end
+      end.uniq
+    rescue StandardError => e
+      warn("Warning: failed to collect Bundler dependency paths: #{e.message}")
+      []
+    end
+  end
+end
diff --git a/rust/rubydex-mcp/src/main.rs b/rust/rubydex-mcp/src/main.rs
index 258e2b7df..d557793c8 100644
--- a/rust/rubydex-mcp/src/main.rs
+++ b/rust/rubydex-mcp/src/main.rs
@@ -10,27 +10,24 @@ mod tools;
     version
 )]
 struct Args {
-    #[arg(value_name = "PATH", default_value = ".")]
-    path: String,
+    #[arg(value_name = "PATH")]
+    paths: Vec<String>,
 }
 
 fn main() {
     let args = Args::parse();
-
-    let root = match std::fs::canonicalize(&args.path) {
-        Ok(p) => p
-            .into_os_string()
-            .into_string()
-            .expect("Project path is not valid UTF-8"),
-        Err(e) => {
-            eprintln!("Warning: failed to canonicalize '{}': {e}", args.path);
-            args.path
-        }
+    let paths = if args.paths.is_empty() {
+        vec![".".to_string()]
+    } else {
+        args.paths
     };
 
+    let paths: Vec<String> = paths.into_iter().map(canonicalize_path).collect();
+    let root = paths.first().expect("expected at least one path").clone();
+
     // Create the server and start indexing in the background.
     let server = server::RubydexServer::new(root.clone());
-    server.spawn_indexer(root);
+    server.spawn_indexer(paths);
 
     // Serve MCP over stdio immediately while indexing runs.
     // We need to do this because Claude Code's default MCP server timeout is 30 seconds,
@@ -46,3 +43,16 @@ fn main() {
         std::process::exit(1);
     }
 }
+
+fn canonicalize_path(path: String) -> String {
+    match std::fs::canonicalize(&path) {
+        Ok(p) => p.into_os_string().into_string().unwrap_or_else(|_| {
+            eprintln!("Warning: canonicalized path for '{path}' is not valid UTF-8, using original");
+            path
+        }),
+        Err(e) => {
+            eprintln!("Warning: failed to canonicalize '{path}': {e}");
+            path
+        }
+    }
+}
diff --git a/rust/rubydex-mcp/src/server.rs b/rust/rubydex-mcp/src/server.rs
index 4d813d6d0..d59c9cb8a 100644
--- a/rust/rubydex-mcp/src/server.rs
+++ b/rust/rubydex-mcp/src/server.rs
@@ -44,11 +44,15 @@ impl RubydexServer {
     }
 
     /// Spawns a background thread that indexes the codebase and marks the server as ready.
-    pub fn spawn_indexer(&self, path: String) {
+    pub fn spawn_indexer(&self, paths: Vec<String>) {
         let state = Arc::clone(&self.state);
         std::thread::spawn(move || {
             let result = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
-                let (file_paths, errors) = rubydex::listing::collect_file_paths(vec![path], &HashSet::new());
+                // Paths may overlap (e.g. a dependency's require_path nested under the workspace, or the
+                // workspace's own gemspec resolving back to the root). This is harmless: the graph keys
+                // documents by UriId and overwrites on re-index, so a file discovered twice costs only
+                // wasted indexing work, never duplicate declarations.
+                let (file_paths, errors) = rubydex::listing::collect_file_paths(paths, &HashSet::new());
                 for error in &errors {
                     eprintln!("Listing error: {error}");
                 }
diff --git a/rust/rubydex-mcp/tests/mcp.rs b/rust/rubydex-mcp/tests/mcp.rs
index cac8beb14..7a99a9d3d 100644
--- a/rust/rubydex-mcp/tests/mcp.rs
+++ b/rust/rubydex-mcp/tests/mcp.rs
@@ -300,3 +300,94 @@ fn mcp_server_e2e() {
         let _ = child.wait().unwrap();
     });
 }
+
+#[test]
+fn mcp_server_indexes_multiple_roots() {
+    with_context(|context| {
+        context.write("workspace/app.rb", "class WorkspaceThing; end");
+        context.write("dependency/lib/dependency_thing.rb", "class DependencyThing; end");
+
+        let mut child = Command::cargo_bin("rubydex_mcp")
+            .unwrap()
+            .args([
+                context.absolute_path_to("workspace").to_str().unwrap(),
+                context.absolute_path_to("dependency/lib").to_str().unwrap(),
+            ])
+            .stdin(Stdio::piped())
+            .stdout(Stdio::piped())
+            .stderr(Stdio::piped())
+            .spawn()
+            .unwrap();
+
+        let mut stdin = child.stdin.take().unwrap();
+        let stdout = child.stdout.take().unwrap();
+        let mut reader = BufReader::new(stdout);
+
+        initialize_session(&mut stdin, &mut reader);
+
+        let mut request_id = 3;
+        wait_for_indexing_to_complete(&mut stdin, &mut reader, &mut request_id);
+
+        let search_response = call_next_tool(
+            &mut stdin,
+            &mut reader,
+            &mut request_id,
+            "search_declarations",
+            &json!({ "query": "DependencyThing" }),
+        );
+        let results = search_response["results"].as_array().unwrap();
+        let result_names = names_from_entries(results);
+        assert_has_name(&result_names, "DependencyThing", "search results");
+
+        drop(stdin);
+        let _ = child.wait().unwrap();
+    });
+}
+
+#[test]
+fn mcp_server_dedups_overlapping_roots() {
+    with_context(|context| {
+        // `lib/thing.rb` lives under the workspace root. Passing both the root and the nested
+        // `lib` directory makes the file discoverable through two roots, mirroring how a
+        // dependency's require_path (or the workspace's own gemspec) can nest under the root.
+        context.write("lib/thing.rb", "class OverlappingThing; end");
+
+        let mut child = Command::cargo_bin("rubydex_mcp")
+            .unwrap()
+            .args([
+                context.absolute_path().to_str().unwrap(),
+                context.absolute_path_to("lib").to_str().unwrap(),
+            ])
+            .stdin(Stdio::piped())
+            .stdout(Stdio::piped())
+            .stderr(Stdio::piped())
+            .spawn()
+            .unwrap();
+
+        let mut stdin = child.stdin.take().unwrap();
+        let stdout = child.stdout.take().unwrap();
+        let mut reader = BufReader::new(stdout);
+
+        initialize_session(&mut stdin, &mut reader);
+
+        let mut request_id = 3;
+        wait_for_indexing_to_complete(&mut stdin, &mut reader, &mut request_id);
+
+        let decl = call_next_tool(
+            &mut stdin,
+            &mut reader,
+            &mut request_id,
+            "get_declaration",
+            &json!({ "name": "OverlappingThing" }),
+        );
+        let definitions = decl["definitions"].as_array().unwrap();
+        assert_eq!(
+            definitions.len(),
+            1,
+            "Expected the doubly-discovered file to yield exactly one definition, got: {decl}"
+        );
+
+        drop(stdin);
+        let _ = child.wait().unwrap();
+    });
+}
diff --git a/test/mcp_server_bridge_test.rb b/test/mcp_server_bridge_test.rb
new file mode 100644
index 000000000..e1cefe14c
--- /dev/null
+++ b/test/mcp_server_bridge_test.rb
@@ -0,0 +1,52 @@
+# frozen_string_literal: true
+
+require_relative "test_helper"
+
+require "rubydex/mcp_server_bridge"
+
+class MCPServerBridgeTest < Minitest::Test
+  def setup
+    @original_bundle_gemfile = ENV.fetch("BUNDLE_GEMFILE", nil)
+  end
+
+  def teardown
+    if @original_bundle_gemfile.nil?
+      ENV.delete("BUNDLE_GEMFILE")
+    else
+      ENV["BUNDLE_GEMFILE"] = @original_bundle_gemfile
+    end
+  end
+
+  def test_compute_index_paths_include_bundler_dependency_require_paths
+    Rubydex::MCPServerBridge.setup_bundler_context(Dir.pwd)
+
+    paths = Rubydex::MCPServerBridge.compute_index_paths([Dir.pwd])
+    spec = Bundler.load.specs.find { |loaded_spec| loaded_spec.name == "rake" }
+    expected_path = File.join(spec.full_gem_path, "lib")
+
+    assert_includes(paths, Dir.pwd)
+    assert_includes(paths, expected_path)
+  end
+
+  def test_passthrough_detects_help_and_version_flags
+    assert(Rubydex::MCPServerBridge.passthrough?(["-h"]))
+    assert(Rubydex::MCPServerBridge.passthrough?(["--help"]))
+    assert(Rubydex::MCPServerBridge.passthrough?(["-V"]))
+    assert(Rubydex::MCPServerBridge.passthrough?(["--version"]))
+    assert(Rubydex::MCPServerBridge.passthrough?(["/some/path", "--help"]))
+  end
+
+  def test_passthrough_is_false_for_paths
+    refute(Rubydex::MCPServerBridge.passthrough?([]))
+    refute(Rubydex::MCPServerBridge.passthrough?(["/some/path"]))
+    refute(Rubydex::MCPServerBridge.passthrough?(["/some/path", "/another/path"]))
+  end
+
+  def test_compute_index_paths_defaults_to_cwd_when_argv_empty
+    Rubydex::MCPServerBridge.setup_bundler_context(Dir.pwd)
+
+    paths = Rubydex::MCPServerBridge.compute_index_paths([])
+
+    assert_includes(paths, Dir.pwd)
+  end
+end