diff --git a/src/main.zig b/src/main.zig
index d6368ff..aa8332f 100644
--- a/src/main.zig
+++ b/src/main.zig
@@ -3598,6 +3598,44 @@ fn runServe(allocator: std.mem.Allocator, stdout: anytype, db: *cobalt.Db, bind:
         }
         sessionStore = cobalt.SessionStore.init(allocator);
         const ttlNanos: i128 = @intFromFloat(seconds * @as(f64, @floatFromInt(std.time.ns_per_s)));
+        server.enableSessions(&sessionStore, ttlNanos);
+        try stdout.print("sessions enabled: AUTH issues a token valid for {d}s\n", .{seconds});
+
+        if (sessionSweepSeconds) |sweepSeconds| {
+            if (sweepSeconds <= 0) {
+                try stdout.writeAll("error: --session-sweep-interval must be a positive number of seconds\n");
+                std.process.exit(2);
+            }
+            const sweepNanos: u64 = @intFromFloat(sweepSeconds * @as(f64, @floatFromInt(std.time.ns_per_s)));
+            server.enableSessionSweep(sweepNanos);
+            try stdout.print("session sweeper enabled: expired sessions dropped every {d}s\n", .{sweepSeconds});
+        }
+    }
+    defer if (sessionTtlSeconds != null) sessionStore.deinit();
+
+    if (ttlSweepSeconds) |seconds| {
+        if (seconds <= 0) {
+            try stdout.writeAll("error: --ttl-sweep-interval must be a positive number of seconds\n");
+            std.process.exit(2);
+        }
+        const nanos: u64 = @intFromFloat(seconds * @as(f64, @floatFromInt(std.time.ns_per_s)));
+        server.enableTtlSweep(nanos);
+        try stdout.print("ttl sweeper enabled: due keys reaped every {d}s\n", .{seconds});
+    }
+
+    if (collTtlSweepSeconds) |seconds| {
+        if (seconds <= 0) {
+            try stdout.writeAll("error: --coll-ttl-sweep-interval must be a positive number of seconds\n");
+            std.process.exit(2);
+        }
+        const nanos: u64 = @intFromFloat(seconds * @as(f64, @floatFromInt(std.time.ns_per_s)));
+        server.enableCollTtlSweep(nanos);
+        try stdout.print("collection ttl sweeper enabled: due collections reaped every {d}s\n", .{seconds});
+    }
+
+    try stdout.print("cobalt server listening on {}\n", .{server.boundAddress()});
+    try server.start();
+    server.wait();
 }
 
 fn runVerify(allocator: std.mem.Allocator, stdout: anytype, dir: []const u8, count: usize) !void {
diff --git a/src/metrics.zig b/src/metrics.zig
new file mode 100644
index 0000000..6507b0d
--- /dev/null
+++ b/src/metrics.zig
@@ -0,0 +1,162 @@
+const std = @import("std");
+const stats_mod = @import("stats.zig");
+
+// Metrics primitives (counter/gauge/histogram), two independent rate-limiter
+// algorithms (token bucket, sliding window counter), and a text-exposition
+// registry. Fully additive module: does not touch Db/txn/storage internals,
+// only depends on stats.zig's public Stats type for the export helper.
+
+pub const MetricCounter = struct {
+    value: std.atomic.Value(u64) = std.atomic.Value(u64).init(0),
+
+    pub fn incr(self: *MetricCounter) void {
+        self.incrBy(1);
+    }
+
+    pub fn incrBy(self: *MetricCounter, delta: u64) void {
+        _ = self.value.fetchAdd(delta, .monotonic);
+    }
+
+    pub fn get(self: *const MetricCounter) u64 {
+        return self.value.load(.monotonic);
+    }
+
+    pub fn reset(self: *MetricCounter) void {
+        self.value.store(0, .monotonic);
+    }
+};
+
+pub const Gauge = struct {
+    bits: std.atomic.Value(u64) = std.atomic.Value(u64).init(0),
+
+    pub fn set(self: *Gauge, v: f64) void {
+        self.bits.store(@bitCast(v), .monotonic);
+    }
+
+    pub fn get(self: *const Gauge) f64 {
+        const raw = self.bits.load(.monotonic);
+        return @bitCast(raw);
+    }
+
+    pub fn add(self: *Gauge, delta: f64) void {
+        while (true) {
+            const cur = self.bits.load(.monotonic);
+            const curF: f64 = @bitCast(cur);
+            const newBits: u64 = @bitCast(curF + delta);
+            if (self.bits.cmpxchgWeak(cur, newBits, .monotonic, .monotonic) == null) return;
+        }
+    }
+};
+
+pub const HistogramError = error{EmptyBounds};
+
+/// Fixed-bucket histogram with Prometheus-style cumulative-bucket quantile
+/// estimation. `bounds` must be strictly increasing and is borrowed (caller
+/// keeps it alive for the histogram's lifetime); values greater than the
+/// last bound land in an implicit +Inf bucket.
+pub const Histogram = struct {
+    bounds: []const f64,
+    counts: []u64,
+    sum: f64 = 0,
+    total: u64 = 0,
+    mutex: std.Thread.Mutex = .{},
+    allocator: std.mem.Allocator,
+
+    pub fn init(allocator: std.mem.Allocator, bounds: []const f64) !Histogram {
+        if (bounds.len == 0) return HistogramError.EmptyBounds;
+        var i: usize = 1;
+        while (i < bounds.len) : (i += 1) {
+            std.debug.assert(bounds[i] > bounds[i - 1]);
+        }
+        const counts = try allocator.alloc(u64, bounds.len + 1);
+        @memset(counts, 0);
+        return .{ .bounds = bounds, .counts = counts, .allocator = allocator };
+    }
+
+    pub fn deinit(self: *Histogram) void {
+        self.allocator.free(self.counts);
+        self.* = undefined;
+    }
+
+    pub fn observe(self: *Histogram, value: f64) void {
+        self.mutex.lock();
+        defer self.mutex.unlock();
+        var idx: usize = self.bounds.len;
+        for (self.bounds, 0..) |bound, i| {
+            if (value <= bound) {
+                idx = i;
+                break;
+            }
+        }
+        self.counts[idx] += 1;
+        self.sum += value;
+        self.total += 1;
+    }
+
+    pub fn count(self: *const Histogram) u64 {
+        return self.total;
+    }
+
+    pub fn mean(self: *const Histogram) f64 {
+        if (self.total == 0) return 0;
+        return self.sum / @as(f64, @floatFromInt(self.total));
+    }
+
+    /// Linear-interpolation quantile estimate over cumulative bucket counts,
+    /// same algorithm Prometheus uses for `histogram_quantile`. q in [0, 1].
+    pub fn quantile(self: *Histogram, q: f64) f64 {
+        self.mutex.lock();
+        defer self.mutex.unlock();
+        if (self.total == 0 or self.bounds.len == 0) return 0;
+        const clamped = std.math.clamp(q, 0, 1);
+        const target = clamped * @as(f64, @floatFromInt(self.total));
+
+        var cumulative: u64 = 0;
+        var i: usize = 0;
+        while (i < self.counts.len) : (i += 1) {
+            const prevCumulative = cumulative;
+            cumulative += self.counts[i];
+            if (@as(f64, @floatFromInt(cumulative)) >= target) {
+                if (i == 0) return self.bounds[0];
+                if (i == self.bounds.len) return self.bounds[self.bounds.len - 1];
+                const lower = self.bounds[i - 1];
+                const upper = self.bounds[i];
+                const countInBucket = cumulative - prevCumulative;
+                if (countInBucket == 0) return upper;
+                const fraction = (target - @as(f64, @floatFromInt(prevCumulative))) / @as(f64, @floatFromInt(countInBucket));
+                return lower + fraction * (upper - lower);
+            }
+        }
+        return self.bounds[self.bounds.len - 1];
+    }
+
+    pub fn bucketCount(self: *Histogram, idx: usize) u64 {
+        self.mutex.lock();
+        defer self.mutex.unlock();
+        return self.counts[idx];
+    }
+};
+
+pub fn linearBuckets(allocator: std.mem.Allocator, start: f64, width: f64, n: usize) ![]f64 {
+    std.debug.assert(n > 0);
+    std.debug.assert(width > 0);
+    const out = try allocator.alloc(f64, n);
+    var i: usize = 0;
+    while (i < n) : (i += 1) out[i] = start + width * @as(f64, @floatFromInt(i));
+    return out;
+}
+
+pub fn exponentialBuckets(allocator: std.mem.Allocator, start: f64, factor: f64, n: usize) ![]f64 {
+    std.debug.assert(n > 0);
+    std.debug.assert(start > 0);
+    std.debug.assert(factor > 1);
+    const out = try allocator.alloc(f64, n);
+    var cur = start;
+    var i: usize = 0;
+    while (i < n) : (i += 1) {
+        out[i] = cur;
+        cur *= factor;
+    }
+    return out;
+}
+