[agent-network] Management: store, manager, synthesizer, policy engine, provider catalog, HTTP/gRPC API

Adds the account-scoped agent-network module: provider/policy/budget CRUD and
store, the reverse-proxy service synthesizer, policy selection + limit
enforcement, the provider catalog (incl. Vertex AI and AWS Bedrock entries),
and the management HTTP + proxy gRPC surfaces.

management/internals/controllers/network_map/controller/repository.go

@@ -3,7 +3,9 @@ package controller
 import (
 	"context"
 
+	"github.com/netbirdio/netbird/management/internals/modules/reverseproxy/service"
 	"github.com/netbirdio/netbird/management/internals/modules/zones"
+	"github.com/netbirdio/netbird/management/server/agentnetwork"
 	"github.com/netbirdio/netbird/management/server/peer"
 	"github.com/netbirdio/netbird/management/server/store"
 	"github.com/netbirdio/netbird/management/server/types"
@@ -16,6 +18,10 @@ type Repository interface {
 	GetPeersByIDs(ctx context.Context, accountID string, peerIDs []string) (map[string]*peer.Peer, error)
 	GetPeerByID(ctx context.Context, accountID string, peerID string) (*peer.Peer, error)
 	GetAccountZones(ctx context.Context, accountID string) ([]*zones.Zone, error)
+	// SynthesizeAgentNetworkServices returns the in-memory reverse-proxy
+	// services synthesised from the account's agent-network provider/policy
+	// state. Empty for accounts without agent-network providers.
+	SynthesizeAgentNetworkServices(ctx context.Context, accountID string) ([]*service.Service, error)
 }
 
 type repository struct {
@@ -50,6 +56,10 @@ func (r *repository) GetPeerByID(ctx context.Context, accountID string, peerID s
 	return r.store.GetPeerByID(ctx, store.LockingStrengthNone, accountID, peerID)
 }
 
+func (r *repository) SynthesizeAgentNetworkServices(ctx context.Context, accountID string) ([]*service.Service, error) {
+	return agentnetwork.SynthesizeServices(ctx, r.store, accountID)
+}
+
 func (r *repository) GetAccountZones(ctx context.Context, accountID string) ([]*zones.Zone, error) {
 	return r.store.GetAccountZones(ctx, store.LockingStrengthNone, accountID)
 }

management/internals/modules/peers/manager.go

@@ -220,12 +220,36 @@ func (m *managerImpl) GetPeerID(ctx context.Context, peerKey string) (string, er
 func (m *managerImpl) CreateProxyPeer(ctx context.Context, accountID string, peerKey string, cluster string) error {
 	existingPeerID, err := m.store.GetPeerIDByKey(ctx, store.LockingStrengthNone, peerKey)
 	if err == nil && existingPeerID != "" {
-		// Peer already exists
+		// Same pubkey already registered — idempotent.
 		return nil
 	}
 
+	// Dedupe stale embedded peer records for the same (account, cluster).
+	// The proxy generates a fresh WireGuard keypair on every startup
+	// (proxy/internal/roundtrip/netbird.go), so without this sweep the
+	// prior embedded peer would linger forever — holding its CGNAT IP
+	// allocation, polluting other peers' rosters, and (most visibly)
+	// leaving the synth DNS pointing at the dead address. The
+	// (account, cluster) tuple identifies "the embedded peer for this
+	// proxy instance at this cluster"; any record matching that tuple
+	// with a different pubkey is by definition stale and must go.
+	staleIDs, err := m.findStaleEmbeddedProxyPeers(ctx, accountID, cluster, peerKey)
+	if err != nil {
+		return fmt.Errorf("scan for stale embedded proxy peers: %w", err)
+	}
+	if len(staleIDs) > 0 {
+		// userID="" + checkConnected=false: the deletion is initiated
+		// by management itself on behalf of the freshly-registering
+		// proxy, not by an end user; the stale peer may still be
+		// marked Connected from its prior session, but its session is
+		// dead by definition (its key no longer exists).
+		if err := m.DeletePeers(ctx, accountID, staleIDs, "", false); err != nil {
+			return fmt.Errorf("delete stale embedded proxy peers %v: %w", staleIDs, err)
+		}
+	}
+
 	name := fmt.Sprintf("proxy-%s", xid.New().String())
-	peer := &peer.Peer{
+	newPeer := &peer.Peer{
 		Ephemeral: true,
 		ProxyMeta: peer.ProxyMeta{
 			Cluster:  cluster,
@@ -242,10 +266,36 @@ func (m *managerImpl) CreateProxyPeer(ctx context.Context, accountID string, pee
 		},
 	}
 
-	_, _, _, _, err = m.accountManager.AddPeer(ctx, accountID, "", "", peer, true)
+	_, _, _, _, err = m.accountManager.AddPeer(ctx, accountID, "", "", newPeer, true)
 	if err != nil {
 		return fmt.Errorf("failed to create proxy peer: %w", err)
 	}
 
 	return nil
 }
+
+// findStaleEmbeddedProxyPeers returns the peer IDs of embedded proxy peer
+// records in accountID that target the same cluster but carry a different
+// WireGuard pubkey than the freshly-registering one. Used by CreateProxyPeer
+// to garbage-collect stale records left behind when the proxy restarts with a
+// regenerated keypair.
+func (m *managerImpl) findStaleEmbeddedProxyPeers(ctx context.Context, accountID, cluster, newKey string) ([]string, error) {
+	account, err := m.store.GetAccount(ctx, accountID)
+	if err != nil {
+		return nil, err
+	}
+	var stale []string
+	for _, p := range account.Peers {
+		if p == nil || !p.ProxyMeta.Embedded {
+			continue
+		}
+		if p.ProxyMeta.Cluster != cluster {
+			continue
+		}
+		if p.Key == newKey {
+			continue
+		}
+		stale = append(stale, p.ID)
+	}
+	return stale, nil
+}

management/internals/modules/reverseproxy/accesslogs/accesslogentry.go

@@ -39,6 +39,10 @@ type AccessLogEntry struct {
 	BytesDownload   int64             `gorm:"index"`
 	Protocol        AccessLogProtocol `gorm:"index"`
 	Metadata        map[string]string `gorm:"serializer:json"`
+	// AgentNetwork marks the entry as emitted by a synthesised agent-network
+	// service. Sourced from proto.AccessLog.AgentNetwork the proxy stamps
+	// before shipping. Indexed so the agent-network log surface filters cheaply.
+	AgentNetwork bool `gorm:"index"`
 }
 
 // FromProto creates an AccessLogEntry from a proto.AccessLog
@@ -58,6 +62,7 @@ func (a *AccessLogEntry) FromProto(serviceLog *proto.AccessLog) {
 	a.BytesDownload = serviceLog.GetBytesDownload()
 	a.Protocol = AccessLogProtocol(serviceLog.GetProtocol())
 	a.Metadata = maps.Clone(serviceLog.GetMetadata())
+	a.AgentNetwork = serviceLog.GetAgentNetwork()
 
 	if sourceIP := serviceLog.GetSourceIp(); sourceIP != "" {
 		if addr, err := netip.ParseAddr(sourceIP); err == nil {

management/internals/modules/reverseproxy/accesslogs/manager/manager.go

@@ -2,12 +2,15 @@ package manager
 
 import (
 	"context"
+	"math"
+	"strconv"
 	"strings"
 	"time"
 
 	log "github.com/sirupsen/logrus"
 
 	"github.com/netbirdio/netbird/management/internals/modules/reverseproxy/accesslogs"
+	agentNetworkTypes "github.com/netbirdio/netbird/management/server/agentnetwork/types"
 	"github.com/netbirdio/netbird/management/server/geolocation"
 	"github.com/netbirdio/netbird/management/server/permissions"
 	"github.com/netbirdio/netbird/management/server/permissions/modules"
@@ -16,6 +19,28 @@ import (
 	"github.com/netbirdio/netbird/shared/management/status"
 )
 
+// Metadata keys the proxy stamps on agent-network access-log entries. These
+// mirror the constants in proxy/internal/middleware/keys.go and form the wire
+// contract between the proxy and management; management flattens them into
+// queryable columns. Keep in sync with the proxy side.
+const (
+	metaKeyProvider           = "llm.provider"
+	metaKeyModel              = "llm.model"
+	metaKeyResolvedProviderID = "llm.resolved_provider_id"
+	metaKeySelectedPolicyID   = "llm.selected_policy_id"
+	metaKeyPolicyDecision     = "llm_policy.decision"
+	metaKeyPolicyReason       = "llm_policy.reason"
+	metaKeyInputTokens        = "llm.input_tokens"  //nolint:gosec // metadata key name, not a credential
+	metaKeyOutputTokens       = "llm.output_tokens" //nolint:gosec // metadata key name, not a credential
+	metaKeyTotalTokens        = "llm.total_tokens"  //nolint:gosec // metadata key name, not a credential
+	metaKeyCostUSDTotal       = "cost.usd_total"
+	metaKeyStream             = "llm.stream"
+	metaKeySessionID          = "llm.session_id"
+	metaKeyAuthorisingGroups  = "llm.authorising_groups"
+	metaKeyRequestPrompt      = "llm.request_prompt"
+	metaKeyResponseCompletion = "llm.response_completion"
+)
+
 type managerImpl struct {
 	store              store.Store
 	permissionsManager permissions.Manager
@@ -31,8 +56,14 @@ func NewManager(store store.Store, permissionsManager permissions.Manager, geo g
 	}
 }
 
-// SaveAccessLog saves an access log entry to the database after enriching it
+// SaveAccessLog saves an access log entry to the database after enriching it.
+// Agent-network entries are flattened into their own dedicated table (queryable
+// LLM columns + group child rows) instead of the shared reverse-proxy table.
 func (m *managerImpl) SaveAccessLog(ctx context.Context, logEntry *accesslogs.AccessLogEntry) error {
+	if logEntry.AgentNetwork {
+		return m.saveAgentNetworkAccessLog(ctx, logEntry)
+	}
+
 	if m.geo != nil && logEntry.GeoLocation.ConnectionIP != nil {
 		location, err := m.geo.Lookup(logEntry.GeoLocation.ConnectionIP)
 		if err != nil {
@@ -61,6 +92,184 @@ func (m *managerImpl) SaveAccessLog(ctx context.Context, logEntry *accesslogs.Ac
 	return nil
 }
 
+// saveAgentNetworkAccessLog flattens the metadata-bearing access-log entry and
+// persists it in two parts:
+//
+//   - The stripped usage record is written unconditionally — usage/cost is
+//     collected on every request regardless of the account's log-collection
+//     toggle (the proxy ships a usage-only entry when logging is disabled).
+//   - The full access-log row (with request detail + prompt) is written only
+//     when the account's EnableLogCollection setting is on. This setting read
+//     is the authoritative gate; the proxy-side strip is defense in depth.
+func (m *managerImpl) saveAgentNetworkAccessLog(ctx context.Context, logEntry *accesslogs.AccessLogEntry) error {
+	entry, groups := flattenAgentNetworkLog(logEntry)
+
+	usage, usageGroups := usageFromFlattenedLog(entry, groups)
+	if err := m.store.CreateAgentNetworkUsage(ctx, usage, usageGroups); err != nil {
+		log.WithContext(ctx).WithFields(log.Fields{
+			"account_id": entry.AccountID,
+			"model":      entry.Model,
+		}).Errorf("failed to save agent-network usage: %v", err)
+		return err
+	}
+
+	settings, err := m.store.GetAgentNetworkSettings(ctx, store.LockingStrengthNone, entry.AccountID)
+	if err != nil {
+		// No settings row (or a transient read error) means we can't confirm
+		// log collection is enabled — usage is already saved, so skip the full
+		// row rather than fail the whole ingest.
+		log.WithContext(ctx).Debugf("skipping full agent-network access-log row for account %s: %v", entry.AccountID, err)
+		return nil
+	}
+	if !settings.EnableLogCollection {
+		return nil
+	}
+
+	if err := m.store.CreateAgentNetworkAccessLog(ctx, entry, groups); err != nil {
+		log.WithContext(ctx).WithFields(log.Fields{
+			"account_id": entry.AccountID,
+			"service_id": entry.ServiceID,
+			"model":      entry.Model,
+			"status":     entry.StatusCode,
+		}).Errorf("failed to save agent-network access log: %v", err)
+		return err
+	}
+	return nil
+}
+
+// flattenAgentNetworkLog converts a reverse-proxy AccessLogEntry (whose LLM
+// dimensions live in the opaque Metadata map) into the flattened
+// agent-network row + authorising-group child rows.
+func flattenAgentNetworkLog(e *accesslogs.AccessLogEntry) (*agentNetworkTypes.AgentNetworkAccessLog, []agentNetworkTypes.AgentNetworkAccessLogGroup) {
+	meta := e.Metadata
+
+	var sourceIP string
+	if e.GeoLocation.ConnectionIP != nil {
+		sourceIP = e.GeoLocation.ConnectionIP.String()
+	}
+
+	entry := &agentNetworkTypes.AgentNetworkAccessLog{
+		ID:            e.ID,
+		AccountID:     e.AccountID,
+		ServiceID:     e.ServiceID,
+		Timestamp:     e.Timestamp,
+		UserID:        e.UserId,
+		SourceIP:      sourceIP,
+		Method:        e.Method,
+		Host:          e.Host,
+		Path:          e.Path,
+		Duration:      e.Duration,
+		StatusCode:    e.StatusCode,
+		AuthMethod:    e.AuthMethodUsed,
+		BytesUpload:   e.BytesUpload,
+		BytesDownload: e.BytesDownload,
+
+		Provider:           meta[metaKeyProvider],
+		Model:              meta[metaKeyModel],
+		SessionID:          meta[metaKeySessionID],
+		ResolvedProviderID: meta[metaKeyResolvedProviderID],
+		SelectedPolicyID:   meta[metaKeySelectedPolicyID],
+		Decision:           meta[metaKeyPolicyDecision],
+		DenyReason:         meta[metaKeyPolicyReason],
+		InputTokens:        parseMetaInt(meta, metaKeyInputTokens),
+		OutputTokens:       parseMetaInt(meta, metaKeyOutputTokens),
+		TotalTokens:        parseMetaInt(meta, metaKeyTotalTokens),
+		CostUSD:            parseMetaFloat(meta, metaKeyCostUSDTotal),
+		Stream:             parseMetaBool(meta, metaKeyStream),
+		RequestPrompt:      meta[metaKeyRequestPrompt],
+		ResponseCompletion: meta[metaKeyResponseCompletion],
+	}
+
+	var groups []agentNetworkTypes.AgentNetworkAccessLogGroup
+	for _, gid := range parseGroupCSV(meta[metaKeyAuthorisingGroups]) {
+		groups = append(groups, agentNetworkTypes.AgentNetworkAccessLogGroup{
+			LogID:     entry.ID,
+			GroupID:   gid,
+			AccountID: entry.AccountID,
+		})
+	}
+	return entry, groups
+}
+
+// usageFromFlattenedLog derives the stripped usage record (and its group child
+// rows) from an already-flattened access-log entry. The usage row shares the
+// log's ID so the two correlate.
+func usageFromFlattenedLog(e *agentNetworkTypes.AgentNetworkAccessLog, groups []agentNetworkTypes.AgentNetworkAccessLogGroup) (*agentNetworkTypes.AgentNetworkUsage, []agentNetworkTypes.AgentNetworkUsageGroup) {
+	usage := &agentNetworkTypes.AgentNetworkUsage{
+		ID:                 e.ID,
+		AccountID:          e.AccountID,
+		Timestamp:          e.Timestamp,
+		UserID:             e.UserID,
+		ResolvedProviderID: e.ResolvedProviderID,
+		Provider:           e.Provider,
+		Model:              e.Model,
+		SessionID:          e.SessionID,
+		InputTokens:        e.InputTokens,
+		OutputTokens:       e.OutputTokens,
+		TotalTokens:        e.TotalTokens,
+		CostUSD:            e.CostUSD,
+	}
+
+	usageGroups := make([]agentNetworkTypes.AgentNetworkUsageGroup, 0, len(groups))
+	for _, g := range groups {
+		usageGroups = append(usageGroups, agentNetworkTypes.AgentNetworkUsageGroup{
+			UsageID:   usage.ID,
+			GroupID:   g.GroupID,
+			AccountID: g.AccountID,
+		})
+	}
+	return usage, usageGroups
+}
+
+// parseMetaInt parses a non-negative token count. Negative or unparseable
+// values are clamped to 0 so a malformed metric can't persist a negative
+// counter.
+func parseMetaInt(meta map[string]string, key string) int64 {
+	if v, err := strconv.ParseInt(strings.TrimSpace(meta[key]), 10, 64); err == nil && v >= 0 {
+		return v
+	}
+	return 0
+}
+
+// parseMetaFloat parses a non-negative, finite cost. Negative, NaN, Inf, or
+// unparseable values are clamped to 0 so a malformed metric can't poison the
+// stored cost.
+func parseMetaFloat(meta map[string]string, key string) float64 {
+	if v, err := strconv.ParseFloat(strings.TrimSpace(meta[key]), 64); err == nil && v >= 0 && !math.IsInf(v, 0) {
+		return v
+	}
+	return 0
+}
+
+func parseMetaBool(meta map[string]string, key string) bool {
+	v, _ := strconv.ParseBool(strings.TrimSpace(meta[key]))
+	return v
+}
+
+// parseGroupCSV splits the comma-separated authorising-group id list the proxy
+// emits, trimming blanks and de-duplicating. Dedup matters because the group
+// rows are keyed by (log_id, group_id) / (usage_id, group_id): a repeated id
+// in the CSV would otherwise produce a duplicate primary key and fail the
+// insert transaction.
+func parseGroupCSV(raw string) []string {
+	if raw == "" {
+		return nil
+	}
+	parts := strings.Split(raw, ",")
+	out := make([]string, 0, len(parts))
+	seen := make(map[string]struct{}, len(parts))
+	for _, p := range parts {
+		if p = strings.TrimSpace(p); p != "" {
+			if _, dup := seen[p]; dup {
+				continue
+			}
+			seen[p] = struct{}{}
+			out = append(out, p)
+		}
+	}
+	return out
+}
+
 // GetAllAccessLogs retrieves access logs for an account with pagination and filtering
 func (m *managerImpl) GetAllAccessLogs(ctx context.Context, accountID, userID string, filter *accesslogs.AccessLogFilter) ([]*accesslogs.AccessLogEntry, int64, error) {
 	ok, ctx, err := m.permissionsManager.ValidateUserPermissions(ctx, accountID, userID, modules.Services, operations.Read)

management/internals/modules/reverseproxy/service/service.go

@@ -66,6 +66,51 @@ type TargetOptions struct {
 	// reachable without WireGuard (public APIs, LAN services, localhost
 	// sidecars). Default false.
 	DirectUpstream bool `json:"direct_upstream,omitempty"`
+	// Middlewares carries per-target agent-network middleware configs. Empty
+	// for private and operator-defined services; populated only by the
+	// agent-network synthesizer.
+	Middlewares             []MiddlewareConfig `gorm:"serializer:json" json:"middlewares,omitempty"`
+	CaptureMaxRequestBytes  int64              `json:"capture_max_request_bytes,omitempty"`
+	CaptureMaxResponseBytes int64              `json:"capture_max_response_bytes,omitempty"`
+	CaptureContentTypes     []string           `gorm:"serializer:json" json:"capture_content_types,omitempty"`
+	// AgentNetwork marks targets synthesised from Agent Network state. The
+	// proxy uses it to gate agent-network-specific behaviour (access log
+	// tagging, observability, etc.).
+	AgentNetwork bool `json:"agent_network,omitempty"`
+	// DisableAccessLog suppresses the per-request access-log emission for this
+	// target. Defaults false to preserve access-log behaviour for every
+	// non-agent-network target. The agent-network synthesizer sets this true
+	// only when the account's EnableLogCollection toggle is off.
+	DisableAccessLog bool `json:"disable_access_log,omitempty"`
+}
+
+// MiddlewareSlot mirrors proto.MiddlewareSlot / middleware.Slot.
+type MiddlewareSlot string
+
+const (
+	MiddlewareSlotOnRequest  MiddlewareSlot = "on_request"
+	MiddlewareSlotOnResponse MiddlewareSlot = "on_response"
+	MiddlewareSlotTerminal   MiddlewareSlot = "terminal"
+)
+
+// MiddlewareFailMode mirrors proto.MiddlewareConfig_FailMode.
+type MiddlewareFailMode string
+
+const (
+	MiddlewareFailOpen   MiddlewareFailMode = "fail_open"
+	MiddlewareFailClosed MiddlewareFailMode = "fail_closed"
+)
+
+// MiddlewareConfig is the per-target configuration for a single
+// middleware instance. Mirrors proto.MiddlewareConfig.
+type MiddlewareConfig struct {
+	ID         string             `json:"id"`
+	Enabled    bool               `json:"enabled"`
+	Slot       MiddlewareSlot     `json:"slot"`
+	ConfigJSON []byte             `json:"config_json,omitempty"`
+	FailMode   MiddlewareFailMode `json:"fail_mode,omitempty"`
+	TimeoutMs  int32              `json:"timeout_ms,omitempty"`
+	CanMutate  bool               `json:"can_mutate"`
 }
 
 type Target struct {
@@ -504,21 +549,75 @@ func targetOptionsToAPI(opts TargetOptions) *api.ServiceTargetOptions {
 
 func targetOptionsToProto(opts TargetOptions) *proto.PathTargetOptions {
 	if !opts.SkipTLSVerify && opts.PathRewrite == "" && opts.RequestTimeout == 0 &&
-		len(opts.CustomHeaders) == 0 && !opts.DirectUpstream {
+		len(opts.CustomHeaders) == 0 && !opts.DirectUpstream &&
+		len(opts.Middlewares) == 0 && opts.CaptureMaxRequestBytes == 0 &&
+		opts.CaptureMaxResponseBytes == 0 && len(opts.CaptureContentTypes) == 0 &&
+		!opts.AgentNetwork && !opts.DisableAccessLog {
 		return nil
 	}
 	popts := &proto.PathTargetOptions{
-		SkipTlsVerify:  opts.SkipTLSVerify,
-		PathRewrite:    pathRewriteToProto(opts.PathRewrite),
-		CustomHeaders:  opts.CustomHeaders,
-		DirectUpstream: opts.DirectUpstream,
+		SkipTlsVerify:    opts.SkipTLSVerify,
+		PathRewrite:      pathRewriteToProto(opts.PathRewrite),
+		CustomHeaders:    opts.CustomHeaders,
+		DirectUpstream:   opts.DirectUpstream,
+		AgentNetwork:     opts.AgentNetwork,
+		DisableAccessLog: opts.DisableAccessLog,
 	}
 	if opts.RequestTimeout != 0 {
 		popts.RequestTimeout = durationpb.New(opts.RequestTimeout)
 	}
+	if len(opts.Middlewares) > 0 {
+		popts.Middlewares = middlewaresToProto(opts.Middlewares)
+	}
+	popts.CaptureMaxRequestBytes = opts.CaptureMaxRequestBytes
+	popts.CaptureMaxResponseBytes = opts.CaptureMaxResponseBytes
+	if len(opts.CaptureContentTypes) > 0 {
+		popts.CaptureContentTypes = append([]string(nil), opts.CaptureContentTypes...)
+	}
 	return popts
 }
 
+// middlewaresToProto converts the internal middleware slice to the proto
+// representation sent to the proxy via the mapping stream.
+func middlewaresToProto(in []MiddlewareConfig) []*proto.MiddlewareConfig {
+	out := make([]*proto.MiddlewareConfig, 0, len(in))
+	for _, m := range in {
+		pm := &proto.MiddlewareConfig{
+			Id:         m.ID,
+			Enabled:    m.Enabled,
+			Slot:       middlewareSlotToProto(m.Slot),
+			ConfigJson: append([]byte(nil), m.ConfigJSON...),
+			CanMutate:  m.CanMutate,
+			FailMode:   middlewareFailModeToProto(m.FailMode),
+		}
+		if m.TimeoutMs > 0 {
+			pm.Timeout = durationpb.New(time.Duration(m.TimeoutMs) * time.Millisecond)
+		}
+		out = append(out, pm)
+	}
+	return out
+}
+
+func middlewareSlotToProto(s MiddlewareSlot) proto.MiddlewareSlot {
+	switch s {
+	case MiddlewareSlotOnRequest:
+		return proto.MiddlewareSlot_MIDDLEWARE_SLOT_ON_REQUEST
+	case MiddlewareSlotOnResponse:
+		return proto.MiddlewareSlot_MIDDLEWARE_SLOT_ON_RESPONSE
+	case MiddlewareSlotTerminal:
+		return proto.MiddlewareSlot_MIDDLEWARE_SLOT_TERMINAL
+	default:
+		return proto.MiddlewareSlot_MIDDLEWARE_SLOT_UNSPECIFIED
+	}
+}
+
+func middlewareFailModeToProto(m MiddlewareFailMode) proto.MiddlewareConfig_FailMode {
+	if m == MiddlewareFailClosed {
+		return proto.MiddlewareConfig_FAIL_CLOSED
+	}
+	return proto.MiddlewareConfig_FAIL_OPEN
+}
+
 // l4TargetOptionsToProto converts L4-relevant target options to proto.
 func l4TargetOptionsToProto(target *Target) *proto.PathTargetOptions {
 	if !target.ProxyProtocol && target.Options.RequestTimeout == 0 && target.Options.SessionIdleTimeout == 0 {

management/internals/server/boot.go

@@ -26,9 +26,11 @@ import (
 	"github.com/netbirdio/netbird/formatter/hook"
 	"github.com/netbirdio/netbird/management/internals/modules/reverseproxy/accesslogs"
 	accesslogsmanager "github.com/netbirdio/netbird/management/internals/modules/reverseproxy/accesslogs/manager"
+	rpservice "github.com/netbirdio/netbird/management/internals/modules/reverseproxy/service"
 	nbgrpc "github.com/netbirdio/netbird/management/internals/shared/grpc"
 	"github.com/netbirdio/netbird/management/server/activity"
 	activitystore "github.com/netbirdio/netbird/management/server/activity/store"
+	"github.com/netbirdio/netbird/management/server/agentnetwork"
 	nbcache "github.com/netbirdio/netbird/management/server/cache"
 	nbContext "github.com/netbirdio/netbird/management/server/context"
 	nbhttp "github.com/netbirdio/netbird/management/server/http"
@@ -120,7 +122,7 @@ func (s *BaseServer) EventStore() activity.Store {
 
 func (s *BaseServer) APIHandler() http.Handler {
 	return Create(s, func() http.Handler {
-		httpAPIHandler, err := nbhttp.NewAPIHandler(context.Background(), s.Router(), s.AccountManager(), s.NetworksManager(), s.ResourcesManager(), s.RoutesManager(), s.GroupsManager(), s.GeoLocationManager(), s.AuthManager(), s.Metrics(), s.PermissionsManager(), s.SettingsManager(), s.ZonesManager(), s.RecordsManager(), s.NetworkMapController(), s.IdpManager(), s.ServiceManager(), s.ReverseProxyDomainManager(), s.AccessLogsManager(), s.ReverseProxyGRPCServer(), s.Config.ReverseProxy.TrustedHTTPProxies, s.RateLimiter(), s.IsValidChildAccount)
+		httpAPIHandler, err := nbhttp.NewAPIHandler(context.Background(), s.Router(), s.AccountManager(), s.NetworksManager(), s.ResourcesManager(), s.RoutesManager(), s.GroupsManager(), s.GeoLocationManager(), s.AuthManager(), s.Metrics(), s.PermissionsManager(), s.SettingsManager(), s.ZonesManager(), s.RecordsManager(), s.NetworkMapController(), s.IdpManager(), s.ServiceManager(), s.ReverseProxyDomainManager(), s.AccessLogsManager(), s.ReverseProxyGRPCServer(), s.Config.ReverseProxy.TrustedHTTPProxies, s.RateLimiter(), s.IsValidChildAccount, s.AgentNetworkManager())
 		if err != nil {
 			log.Fatalf("failed to create API handler: %v", err)
 		}
@@ -223,11 +225,35 @@ func (s *BaseServer) ReverseProxyGRPCServer() *nbgrpc.ProxyServiceServer {
 		s.AfterInit(func(s *BaseServer) {
 			proxyService.SetServiceManager(s.ServiceManager())
 			proxyService.SetProxyController(s.ServiceProxyController())
+			proxyService.SetAgentNetworkSynthesizer(newAgentNetworkSynthesizer(s.Store()))
+			proxyService.SetAgentNetworkLimitsService(s.AgentNetworkManager())
 		})
 		return proxyService
 	})
 }
 
+// agentNetworkSynthesizerAdapter implements nbgrpc.AgentNetworkSynthesizer by
+// delegating to the agentnetwork package's store-backed synthesiser.
+type agentNetworkSynthesizerAdapter struct {
+	store store.Store
+}
+
+func newAgentNetworkSynthesizer(s store.Store) *agentNetworkSynthesizerAdapter {
+	return &agentNetworkSynthesizerAdapter{store: s}
+}
+
+func (a *agentNetworkSynthesizerAdapter) SynthesizeServicesForCluster(ctx context.Context, clusterAddr string) ([]*rpservice.Service, error) {
+	return agentnetwork.SynthesizeServicesForCluster(ctx, a.store, clusterAddr)
+}
+
+func (a *agentNetworkSynthesizerAdapter) SynthesizeServicesForAccount(ctx context.Context, accountID string) ([]*rpservice.Service, error) {
+	return agentnetwork.SynthesizeServices(ctx, a.store, accountID)
+}
+
+func (a *agentNetworkSynthesizerAdapter) SynthesizeServiceForDomain(ctx context.Context, domain string) (*rpservice.Service, error) {
+	return agentnetwork.SynthesizeServiceForDomain(ctx, a.store, domain)
+}
+
 func (s *BaseServer) proxyOIDCConfig() nbgrpc.ProxyOIDCConfig {
 	return Create(s, func() nbgrpc.ProxyOIDCConfig {
 		return nbgrpc.ProxyOIDCConfig{

management/internals/server/modules.go

@@ -20,6 +20,7 @@ import (
 	recordsManager "github.com/netbirdio/netbird/management/internals/modules/zones/records/manager"
 	"github.com/netbirdio/netbird/management/server"
 	"github.com/netbirdio/netbird/management/server/account"
+	"github.com/netbirdio/netbird/management/server/agentnetwork"
 	"github.com/netbirdio/netbird/management/server/geolocation"
 	"github.com/netbirdio/netbird/management/server/groups"
 	"github.com/netbirdio/netbird/management/server/idp"
@@ -194,6 +195,24 @@ func (s *BaseServer) NetworksManager() networks.Manager {
 	})
 }
 
+func (s *BaseServer) AgentNetworkManager() agentnetwork.Manager {
+	return Create(s, func() agentnetwork.Manager {
+		mgr := agentnetwork.NewManager(
+			s.Store(),
+			s.PermissionsManager(),
+			s.AccountManager(),
+			s.ServiceProxyController(),
+		)
+		// Sweep expired agent-network access logs per account retention,
+		// reusing the reverse-proxy cleanup interval config.
+		mgr.StartAccessLogCleanup(
+			context.Background(),
+			s.Config.ReverseProxy.AccessLogCleanupIntervalHours,
+		)
+		return mgr
+	})
+}
+
 func (s *BaseServer) ZonesManager() zones.Manager {
 	return Create(s, func() zones.Manager {
 		return zonesManager.NewManager(s.Store(), s.AccountManager(), s.PermissionsManager(), s.DNSDomain())

management/internals/shared/grpc/proxy.go

@@ -10,6 +10,7 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"math"
 	"net"
 	"net/http"
 	"net/url"
@@ -35,6 +36,7 @@ import (
 	"github.com/netbirdio/netbird/management/internals/modules/reverseproxy/sessionkey"
 	"github.com/netbirdio/netbird/management/server/idp"
 	"github.com/netbirdio/netbird/management/server/peer"
+	"github.com/netbirdio/netbird/management/server/agentnetwork"
 	"github.com/netbirdio/netbird/management/server/types"
 	"github.com/netbirdio/netbird/management/server/users"
 	proxyauth "github.com/netbirdio/netbird/proxy/auth"
@@ -60,6 +62,23 @@ type ProxyTokenChecker interface {
 }
 
 // ProxyServiceServer implements the ProxyService gRPC server
+// AgentNetworkSynthesizer produces in-memory reverse-proxy services from
+// Agent Network provider/policy state for the proxy snapshot path; synthesised
+// services never appear in the reverseproxy_services table.
+type AgentNetworkSynthesizer interface {
+	SynthesizeServicesForCluster(ctx context.Context, clusterAddr string) ([]*rpservice.Service, error)
+	SynthesizeServicesForAccount(ctx context.Context, accountID string) ([]*rpservice.Service, error)
+	SynthesizeServiceForDomain(ctx context.Context, domain string) (*rpservice.Service, error)
+}
+
+// AgentNetworkLimitsService is the minimal slice of agentnetwork.Manager the
+// gRPC layer needs for CheckLLMPolicyLimits + RecordLLMUsage — kept narrow so
+// the grpc package doesn't take a hard import on the full manager.
+type AgentNetworkLimitsService interface {
+	SelectPolicyForRequest(ctx context.Context, in agentnetwork.PolicySelectionInput) (*agentnetwork.PolicySelectionResult, error)
+	RecordUsage(ctx context.Context, in agentnetwork.RecordUsageInput) error
+}
+
 type ProxyServiceServer struct {
 	proto.UnimplementedProxyServiceServer
 
@@ -72,6 +91,14 @@ type ProxyServiceServer struct {
 	mu sync.RWMutex
 	// Manager for reverse proxy operations
 	serviceManager rpservice.Manager
+	// agentNetworkSynth produces synthesised reverse-proxy services from
+	// Agent Network state. Optional — when nil the snapshot path only ships
+	// persisted services.
+	agentNetworkSynth AgentNetworkSynthesizer
+	// agentNetworkLimits handles the pre-flight selection (CheckLLMPolicyLimits)
+	// and the post-flight consumption write (RecordLLMUsage). Optional — when
+	// nil both RPCs return Unimplemented.
+	agentNetworkLimits AgentNetworkLimitsService
 	// ProxyController for service updates and cluster management
 	proxyController proxy.Controller
 
@@ -209,6 +236,127 @@ func (s *ProxyServiceServer) SetServiceManager(manager rpservice.Manager) {
 	s.serviceManager = manager
 }
 
+// SetAgentNetworkSynthesizer wires the agent-network service synthesiser.
+// Optional — when nil the snapshot path skips agent-network synthesis. The
+// modules layer injects this after both the proxy server and the agent-network
+// manager are constructed.
+func (s *ProxyServiceServer) SetAgentNetworkSynthesizer(synth AgentNetworkSynthesizer) {
+	s.mu.Lock()
+	s.agentNetworkSynth = synth
+	s.mu.Unlock()
+}
+
+// SetAgentNetworkLimitsService wires the policy-selection + post-flight
+// consumption sink. Pass nil to disable; both RPCs return Unimplemented while
+// unset so partial wiring surfaces during integration.
+func (s *ProxyServiceServer) SetAgentNetworkLimitsService(svc AgentNetworkLimitsService) {
+	s.mu.Lock()
+	s.agentNetworkLimits = svc
+	s.mu.Unlock()
+}
+
+// agentNetworkSynthesizer returns the synthesiser under read lock.
+func (s *ProxyServiceServer) agentNetworkSynthesizer() AgentNetworkSynthesizer {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	return s.agentNetworkSynth
+}
+
+// CheckLLMPolicyLimits is the pre-flight policy gate the proxy calls before
+// forwarding an LLM request upstream. Delegates to the agent-network selector,
+// which scores applicable policies by remaining headroom and returns the
+// policy that pays for this request (or a deny when all are exhausted).
+func (s *ProxyServiceServer) CheckLLMPolicyLimits(ctx context.Context, req *proto.CheckLLMPolicyLimitsRequest) (*proto.CheckLLMPolicyLimitsResponse, error) {
+	s.mu.RLock()
+	svc := s.agentNetworkLimits
+	s.mu.RUnlock()
+	if svc == nil {
+		return nil, status.Errorf(codes.Unimplemented, "agent-network limits service not configured on management")
+	}
+	if req.GetAccountId() == "" {
+		return nil, status.Errorf(codes.InvalidArgument, "account_id is required")
+	}
+	if err := enforceAccountScope(ctx, req.GetAccountId()); err != nil {
+		return nil, err
+	}
+
+	res, err := svc.SelectPolicyForRequest(ctx, agentnetwork.PolicySelectionInput{
+		AccountID:  req.GetAccountId(),
+		UserID:     req.GetUserId(),
+		GroupIDs:   req.GetGroupIds(),
+		ProviderID: req.GetProviderId(),
+	})
+	if err != nil {
+		log.WithContext(ctx).Errorf("select policy for request: %v", err)
+		return nil, status.Error(codes.Internal, "select policy failed")
+	}
+
+	if !res.Allow {
+		return &proto.CheckLLMPolicyLimitsResponse{
+			Decision:           "deny",
+			SelectedPolicyId:   res.SelectedPolicyID,
+			AttributionGroupId: res.AttributionGroupID,
+			WindowSeconds:      res.WindowSeconds,
+			DenyCode:           res.DenyCode,
+			DenyReason:         res.DenyReason,
+		}, nil
+	}
+	return &proto.CheckLLMPolicyLimitsResponse{
+		Decision:           "allow",
+		SelectedPolicyId:   res.SelectedPolicyID,
+		AttributionGroupId: res.AttributionGroupID,
+		WindowSeconds:      res.WindowSeconds,
+	}, nil
+}
+
+// RecordLLMUsage increments the per-(dimension, window) consumption counter for
+// the user and optional attribution group after a served request. Returns
+// Unimplemented when the agent-network limits service hasn't been wired.
+func (s *ProxyServiceServer) RecordLLMUsage(ctx context.Context, req *proto.RecordLLMUsageRequest) (*proto.RecordLLMUsageResponse, error) {
+	s.mu.RLock()
+	svc := s.agentNetworkLimits
+	s.mu.RUnlock()
+	if svc == nil {
+		return nil, status.Errorf(codes.Unimplemented, "agent-network limits service not configured on management")
+	}
+
+	accountID := req.GetAccountId()
+	if accountID == "" {
+		return nil, status.Errorf(codes.InvalidArgument, "account_id is required")
+	}
+	if err := enforceAccountScope(ctx, accountID); err != nil {
+		return nil, err
+	}
+	tokensIn := req.GetTokensInput()
+	tokensOut := req.GetTokensOutput()
+	costUSD := req.GetCostUsd()
+
+	// Reject impossible counters at the boundary instead of recording them:
+	// a negative window, negative tokens, or a negative / non-finite cost
+	// would otherwise decrement or poison the persisted consumption totals.
+	if req.GetWindowSeconds() < 0 || tokensIn < 0 || tokensOut < 0 || costUSD < 0 || math.IsNaN(costUSD) || math.IsInf(costUSD, 0) {
+		return nil, status.Errorf(codes.InvalidArgument, "usage counters must be non-negative and finite")
+	}
+
+	// Book the policy-window dimensions (when a policy cap bound this request)
+	// and every applicable account budget rule's window in a single batched
+	// transaction.
+	if err := svc.RecordUsage(ctx, agentnetwork.RecordUsageInput{
+		AccountID:          accountID,
+		UserID:             req.GetUserId(),
+		AttributionGroupID: req.GetGroupId(),
+		GroupIDs:           req.GetGroupIds(),
+		WindowSeconds:      req.GetWindowSeconds(),
+		TokensIn:           tokensIn,
+		TokensOut:          tokensOut,
+		CostUSD:            costUSD,
+	}); err != nil {
+		log.WithContext(ctx).Errorf("record usage: %v", err)
+		return nil, status.Error(codes.Internal, "record usage failed")
+	}
+	return &proto.RecordLLMUsageResponse{}, nil
+}
+
 // SetProxyController sets the proxy controller. Must be called before serving.
 func (s *ProxyServiceServer) SetProxyController(proxyController proxy.Controller) {
 	s.mu.Lock()
@@ -623,12 +771,40 @@ func (s *ProxyServiceServer) snapshotServiceMappings(ctx context.Context, conn *
 		return nil, fmt.Errorf("get services from store: %w", err)
 	}
 
+	if synth := s.agentNetworkSynthesizer(); synth != nil {
+		var synthesised []*rpservice.Service
+		var serr error
+		// Account-scoped connections synthesise only their own account, so the
+		// snapshot can never carry another tenant's mappings (which embed the
+		// upstream auth header derived from that tenant's provider API key).
+		// Global connections still see the whole cluster.
+		if conn.accountID != nil {
+			synthesised, serr = synth.SynthesizeServicesForAccount(ctx, *conn.accountID)
+		} else {
+			synthesised, serr = synth.SynthesizeServicesForCluster(ctx, conn.address)
+		}
+		if serr != nil {
+			// Surface a real synthesis failure instead of silently shipping an
+			// incomplete snapshot (which would drop the account's agent-network
+			// routes). Consistent with the persisted-services error above; the
+			// proxy retries the snapshot on connection error.
+			return nil, fmt.Errorf("synthesise agent-network services: %w", serr)
+		}
+		services = append(services, synthesised...)
+	}
+
 	oidcCfg := s.GetOIDCValidationConfig()
 	var mappings []*proto.ProxyMapping
 	for _, service := range services {
 		if !service.Enabled || service.ProxyCluster == "" || service.ProxyCluster != conn.address {
 			continue
 		}
+		// Defense in depth: an account-scoped proxy must never receive another
+		// account's mapping, matching the per-account filtering the incremental
+		// update path already applies.
+		if conn.accountID != nil && service.AccountID != *conn.accountID {
+			continue
+		}
 
 		m := service.ToProtoMapping(rpservice.Create, "", oidcCfg)
 		if !proxyAcceptsMapping(conn, m) {
@@ -1617,7 +1793,29 @@ func (s *ProxyServiceServer) ValidateSession(ctx context.Context, req *proto.Val
 }
 
 func (s *ProxyServiceServer) getServiceByDomain(ctx context.Context, domain string) (*rpservice.Service, error) {
-	return s.serviceManager.GetServiceByDomain(ctx, domain)
+	service, err := s.serviceManager.GetServiceByDomain(ctx, domain)
+	if err == nil {
+		return service, nil
+	}
+
+	// Fall back to the Agent Network synthesiser scoped directly to the domain's
+	// account. Synthesised services are never persisted, so they must resolve
+	// here for OIDC / session / tunnel-peer flows against agent-network
+	// endpoints. Resolving by domain synthesises only the owning account rather
+	// than every tenant on the cluster.
+	if synth := s.agentNetworkSynthesizer(); synth != nil {
+		svc, serr := synth.SynthesizeServiceForDomain(ctx, domain)
+		if serr != nil {
+			// A real synthesis failure must surface, not be masked by the
+			// original store miss — otherwise a transient DB error looks like
+			// "no such service".
+			return nil, fmt.Errorf("synthesize agent-network service for %s: %w", domain, serr)
+		}
+		if svc != nil {
+			return svc, nil
+		}
+	}
+	return nil, err
 }
 
 func (s *ProxyServiceServer) checkGroupAccess(service *rpservice.Service, user *types.User) error {