diff --git a/stacks/llama-cpp/main.tf b/stacks/llama-cpp/main.tf
index 289f058f..27c952d6 100644
--- a/stacks/llama-cpp/main.tf
+++ b/stacks/llama-cpp/main.tf
@@ -17,6 +17,8 @@ locals {
   # snapshot_download with allow_patterns). Stable symlinks model.gguf /
   # mmproj.gguf are created after download so llama-swap config can be
   # filename-agnostic.
+  # `text_only = true` skips mmproj download + --mmproj flag (text-only LLM).
+  # Vision models keep `text_only = false` (default).
   models = {
     qwen3vl-8b = {
       hf_repo        = "Qwen/Qwen3-VL-8B-Instruct-GGUF"
@@ -24,6 +26,7 @@ locals {
       mmproj_pattern = "*mmproj*.gguf"
       ctx_size       = 3072
       gpu_layers     = 99
+      text_only      = false
     }
     minicpm-v-4-5 = {
       hf_repo        = "openbmb/MiniCPM-V-4_5-gguf"
@@ -31,6 +34,7 @@ locals {
       mmproj_pattern = "*mmproj*.gguf"
       ctx_size       = 3072
       gpu_layers     = 99
+      text_only      = false
     }
     qwen3vl-4b = {
       hf_repo        = "Qwen/Qwen3-VL-4B-Instruct-GGUF"
@@ -38,6 +42,21 @@ locals {
       mmproj_pattern = "*mmproj*.gguf"
       ctx_size       = 3072
       gpu_layers     = 99
+      text_only      = false
+    }
+    # Text-only triage / drafting model for recruiter-responder.
+    # Q4_K_M, ~4.7GB, 32k native context (capped at 16k here — plenty
+    # for recruiter emails + extraction prompt + JSON output).
+    # Unsloth's GGUF: well-maintained, includes Q4_K_M. Qwen3 is a
+    # thinking-capable model; recruiter-responder disables thinking via
+    # `enable_thinking=false` in the chat-template kwargs.
+    qwen3-8b = {
+      hf_repo        = "unsloth/Qwen3-8B-GGUF"
+      gguf_pattern   = "*Q4_K_M*.gguf"
+      mmproj_pattern = ""
+      ctx_size       = 16384
+      gpu_layers     = 99
+      text_only      = true
     }
   }
 
@@ -55,18 +74,20 @@ locals {
 
     models = {
       for mid, cfg in local.models : mid => {
-        cmd = join(" ", [
+        cmd = join(" ", concat([
           "/app/llama-server",
           "--host 0.0.0.0",
           "--port $${PORT}",
           "-m /models/${mid}/model.gguf",
+          ], cfg.text_only ? [] : [
           "--mmproj /models/${mid}/mmproj.gguf",
+          ], [
           "-ngl ${cfg.gpu_layers}",
           "-c ${cfg.ctx_size}",
           "-np 1",
           "--jinja",
           "-fa on",
-        ])
+        ]))
         ttl           = 600 # unload after 10 min idle
         checkEndpoint = "/health"
       }
@@ -133,11 +154,15 @@ resource "kubernetes_job_v1" "download_models" {
             for mid, cfg in models.items():
                 local_dir = f"/models/{mid}"
                 os.makedirs(local_dir, exist_ok=True)
-                print(f"==> downloading {mid} from {cfg['hf_repo']} -> {local_dir}", flush=True)
+                text_only = cfg.get("text_only", False)
+                patterns = [cfg["gguf_pattern"]]
+                if not text_only and cfg.get("mmproj_pattern"):
+                    patterns.append(cfg["mmproj_pattern"])
+                print(f"==> downloading {mid} from {cfg['hf_repo']} -> {local_dir} (text_only={text_only})", flush=True)
                 snapshot_download(
                     repo_id=cfg["hf_repo"],
                     local_dir=local_dir,
-                    allow_patterns=[cfg["gguf_pattern"], cfg["mmproj_pattern"]],
+                    allow_patterns=patterns,
                     token=os.environ.get("HF_TOKEN") or None,
                     # Single-threaded download — multi-worker buffers
                     # multi-GB chunks per worker and OOMs the Job at 2Gi.
@@ -146,17 +171,20 @@ resource "kubernetes_job_v1" "download_models" {
                 # Resolve actual filenames and create stable symlinks so
                 # llama-swap config is filename-agnostic.
                 ggufs = [p for p in glob.glob(f"{local_dir}/*Q4_K_M*.gguf") if "mmproj" not in p.lower()]
-                mmprojs = glob.glob(f"{local_dir}/*mmproj*.gguf")
                 if not ggufs:
                     raise SystemExit(f"no GGUF found in {local_dir}")
-                if not mmprojs:
-                    raise SystemExit(f"no mmproj found in {local_dir}")
                 gguf_link = f"{local_dir}/model.gguf"
-                mmproj_link = f"{local_dir}/mmproj.gguf"
-                for link, target in ((gguf_link, ggufs[0]), (mmproj_link, mmprojs[0])):
-                    if os.path.islink(link) or os.path.exists(link):
-                        os.unlink(link)
-                    os.symlink(os.path.basename(target), link)
+                if os.path.islink(gguf_link) or os.path.exists(gguf_link):
+                    os.unlink(gguf_link)
+                os.symlink(os.path.basename(ggufs[0]), gguf_link)
+                if not text_only:
+                    mmprojs = glob.glob(f"{local_dir}/*mmproj*.gguf")
+                    if not mmprojs:
+                        raise SystemExit(f"no mmproj found in {local_dir}")
+                    mmproj_link = f"{local_dir}/mmproj.gguf"
+                    if os.path.islink(mmproj_link) or os.path.exists(mmproj_link):
+                        os.unlink(mmproj_link)
+                    os.symlink(os.path.basename(mmprojs[0]), mmproj_link)
                 print(f"==> done {mid}", flush=True)
                 for f in sorted(os.listdir(local_dir)):
                     full = os.path.join(local_dir, f)
diff --git a/stacks/openclaw/main.tf b/stacks/openclaw/main.tf
index 34442301..7bccd0d1 100644
--- a/stacks/openclaw/main.tf
+++ b/stacks/openclaw/main.tf
@@ -456,6 +456,32 @@ resource "kubernetes_deployment" "openclaw" {
         # Dotfiles already exist on NFS at /home/node/.openclaw/dotfiles from
         # a previous clone. To update, run git pull manually or via CronJob.
 
+        # Init 3: install the recruiter-api OpenClaw plugin from the
+        # recruiter-responder image into NFS extensions/. Plugin lifecycle
+        # is coupled to the recruiter-responder image tag — bumping that
+        # tag re-installs the plugin on next openclaw pod restart.
+        init_container {
+          name  = "install-recruiter-plugin"
+          image = "forgejo.viktorbarzin.me/viktor/recruiter-responder:latest"
+          command = ["sh", "-c", <<-EOT
+            set -eu
+            mkdir -p /home/node/.openclaw/extensions/recruiter-api
+            cp -r /app/openclaw-plugin/. /home/node/.openclaw/extensions/recruiter-api/
+            chown -R 1000:1000 /home/node/.openclaw/extensions/recruiter-api
+            echo "recruiter-api plugin installed at /home/node/.openclaw/extensions/recruiter-api"
+            ls -la /home/node/.openclaw/extensions/recruiter-api
+          EOT
+          ]
+          volume_mount {
+            name       = "openclaw-home"
+            mount_path = "/home/node/.openclaw"
+          }
+          resources {
+            requests = { cpu = "50m", memory = "64Mi" }
+            limits   = { memory = "128Mi" }
+          }
+        }
+
         # Main container: OpenClaw
         container {
           name  = "openclaw"
@@ -533,6 +559,23 @@ resource "kubernetes_deployment" "openclaw" {
               }
             }
           }
+          # Recruiter Responder API — consumed by the recruiter-api plugin
+          # (mounted into /home/node/.openclaw/extensions/recruiter-api/ via
+          # the install-recruiter-plugin init container below).
+          env {
+            name  = "RECRUITER_RESPONDER_URL"
+            value = "http://recruiter-responder.recruiter-responder.svc.cluster.local:8080"
+          }
+          env {
+            name = "RECRUITER_RESPONDER_TOKEN"
+            value_from {
+              secret_key_ref {
+                name     = "openclaw-secrets"
+                key      = "recruiter_responder_bearer_token"
+                optional = true
+              }
+            }
+          }
           # Python packages path for skills
           env {
             name  = "PYTHONPATH"
diff --git a/stacks/recruiter-responder/main.tf b/stacks/recruiter-responder/main.tf
new file mode 100644
index 00000000..f8685567
--- /dev/null
+++ b/stacks/recruiter-responder/main.tf
@@ -0,0 +1,302 @@
+variable "image_tag" {
+  type        = string
+  default     = "latest"
+  description = "recruiter-responder image tag. Use 8-char git SHA in CI."
+}
+
+variable "postgresql_host" { type = string }
+
+locals {
+  namespace = "recruiter-responder"
+  image     = "forgejo.viktorbarzin.me/viktor/recruiter-responder:${var.image_tag}"
+  labels = {
+    app = "recruiter-responder"
+  }
+}
+
+resource "kubernetes_namespace" "recruiter_responder" {
+  metadata {
+    name = local.namespace
+    labels = {
+      tier              = local.tiers.aux
+      "istio-injection" = "disabled"
+    }
+  }
+  lifecycle {
+    # KYVERNO_LIFECYCLE_V1: goldilocks-vpa-auto-mode ClusterPolicy stamps this label.
+    ignore_changes = [metadata[0].labels["goldilocks.fairwinds.com/vpa-update-mode"]]
+  }
+}
+
+# App secrets — seed these in Vault before applying:
+#   secret/recruiter-responder
+#     webhook_bearer_token   — bearer for all /api/* endpoints (consumed by
+#                              the OpenClaw recruiter-api plugin)
+#     imap_me_user           — IMAP for me@viktorbarzin.me (full address)
+#     imap_me_pass           — IMAP password for me@
+#     imap_spam_user         — IMAP for spam@viktorbarzin.me
+#     imap_spam_pass         — IMAP password for spam@
+#     smtp_password          — SMTP password for me@viktorbarzin.me
+#     claude_agent_token     — Bearer for claude-agent-service (Tier-2)
+#     task_webhook_token     — Bearer for OpenClaw task-webhook (optional;
+#                              empty allowed if task-webhook is unauthed)
+#
+# Schema in CNPG: `recruiter_responder` (alembic creates on first migrate).
+# DB user: created via Vault database engine — see static-creds/pg-recruiter-responder.
+resource "kubernetes_manifest" "external_secret" {
+  manifest = {
+    apiVersion = "external-secrets.io/v1beta1"
+    kind       = "ExternalSecret"
+    metadata = {
+      name      = "recruiter-responder-secrets"
+      namespace = local.namespace
+    }
+    spec = {
+      refreshInterval = "15m"
+      secretStoreRef = {
+        name = "vault-kv"
+        kind = "ClusterSecretStore"
+      }
+      target = {
+        name = "recruiter-responder-secrets"
+        template = {
+          metadata = {
+            annotations = {
+              "reloader.stakater.com/match" = "true"
+            }
+          }
+        }
+      }
+      data = [
+        { secretKey = "WEBHOOK_BEARER_TOKEN", remoteRef = { key = "recruiter-responder", property = "webhook_bearer_token" } },
+        { secretKey = "IMAP_ME_USER", remoteRef = { key = "recruiter-responder", property = "imap_me_user" } },
+        { secretKey = "IMAP_ME_PASS", remoteRef = { key = "recruiter-responder", property = "imap_me_pass" } },
+        { secretKey = "IMAP_SPAM_USER", remoteRef = { key = "recruiter-responder", property = "imap_spam_user" } },
+        { secretKey = "IMAP_SPAM_PASS", remoteRef = { key = "recruiter-responder", property = "imap_spam_pass" } },
+        { secretKey = "SMTP_PASSWORD", remoteRef = { key = "recruiter-responder", property = "smtp_password" } },
+        { secretKey = "CLAUDE_AGENT_TOKEN", remoteRef = { key = "recruiter-responder", property = "claude_agent_token" } },
+        { secretKey = "TASK_WEBHOOK_TOKEN", remoteRef = { key = "recruiter-responder", property = "task_webhook_token" } },
+      ]
+    }
+  }
+  depends_on = [kubernetes_namespace.recruiter_responder]
+}
+
+# DB credentials from Vault database engine (7-day rotation).
+# Builds the asyncpg DSN consumed by the FastAPI app as DB_CONNECTION_STRING.
+# Pre-req in dbaas: CNPG cluster has DB `recruiter_responder`, role
+# `recruiter_responder`, and Vault role `static-creds/pg-recruiter-responder`.
+resource "kubernetes_manifest" "db_external_secret" {
+  manifest = {
+    apiVersion = "external-secrets.io/v1beta1"
+    kind       = "ExternalSecret"
+    metadata = {
+      name      = "recruiter-responder-db-creds"
+      namespace = local.namespace
+    }
+    spec = {
+      refreshInterval = "15m"
+      secretStoreRef = {
+        name = "vault-database"
+        kind = "ClusterSecretStore"
+      }
+      target = {
+        name = "recruiter-responder-db-creds"
+        template = {
+          metadata = {
+            annotations = {
+              "reloader.stakater.com/match" = "true"
+            }
+          }
+          data = {
+            DB_CONNECTION_STRING = "postgresql+asyncpg://recruiter_responder:{{ .password }}@${var.postgresql_host}:5432/recruiter_responder"
+            DB_PASSWORD          = "{{ .password }}"
+          }
+        }
+      }
+      data = [{
+        secretKey = "password"
+        remoteRef = {
+          key      = "static-creds/pg-recruiter-responder"
+          property = "password"
+        }
+      }]
+    }
+  }
+  depends_on = [kubernetes_namespace.recruiter_responder]
+}
+
+resource "kubernetes_deployment" "recruiter_responder" {
+  metadata {
+    name      = "recruiter-responder"
+    namespace = kubernetes_namespace.recruiter_responder.metadata[0].name
+    labels = merge(local.labels, {
+      tier = local.tiers.aux
+    })
+    annotations = {
+      "reloader.stakater.com/search" = "true"
+    }
+  }
+
+  spec {
+    # IMAP IDLE + APScheduler want a single leader; concurrency hurts both.
+    replicas = 1
+    strategy {
+      type = "Recreate"
+    }
+
+    selector {
+      match_labels = local.labels
+    }
+
+    template {
+      metadata {
+        labels = local.labels
+      }
+
+      spec {
+        image_pull_secrets {
+          name = "registry-credentials"
+        }
+
+        init_container {
+          name    = "alembic-migrate"
+          image   = local.image
+          command = ["alembic", "upgrade", "head"]
+
+          env_from {
+            secret_ref { name = "recruiter-responder-secrets" }
+          }
+          env_from {
+            secret_ref { name = "recruiter-responder-db-creds" }
+          }
+
+          resources {
+            requests = { cpu = "50m", memory = "256Mi" }
+            limits   = { memory = "512Mi" }
+          }
+        }
+
+        container {
+          name  = "recruiter-responder"
+          image = local.image
+
+          port {
+            container_port = 8080
+          }
+
+          env_from {
+            secret_ref { name = "recruiter-responder-secrets" }
+          }
+          env_from {
+            secret_ref { name = "recruiter-responder-db-creds" }
+          }
+
+          # IMAP fan-in: read both mailboxes off the in-cluster mailserver.
+          env {
+            name  = "IMAP_MAILBOXES"
+            value = "me,spam"
+          }
+          env {
+            name  = "IMAP_ME_HOST"
+            value = "mailserver.mailserver.svc.cluster.local"
+          }
+          env {
+            name  = "IMAP_SPAM_HOST"
+            value = "mailserver.mailserver.svc.cluster.local"
+          }
+          # SMTP (outbound reply) — same mailserver service, STARTTLS on 587.
+          env {
+            name  = "SMTP_HOST"
+            value = "mailserver.mailserver.svc.cluster.local"
+          }
+          env {
+            name  = "SMTP_PORT"
+            value = "587"
+          }
+          env {
+            name  = "SMTP_USER"
+            value = "me@viktorbarzin.me"
+          }
+          env {
+            name  = "SMTP_FROM_ADDR"
+            value = "me@viktorbarzin.me"
+          }
+          env {
+            name  = "SMTP_FROM_NAME"
+            value = "Viktor Barzin"
+          }
+          # Tier-0 LLM
+          env {
+            name  = "LLAMA_SWAP_URL"
+            value = "http://llama-swap.llama-cpp.svc.cluster.local:8080"
+          }
+          env {
+            name  = "LLAMA_SWAP_MODEL"
+            value = "qwen3-8b"
+          }
+          # Tier-2 LLM (deep_research only)
+          env {
+            name  = "CLAUDE_AGENT_URL"
+            value = "http://claude-agent-service.claude-agent.svc.cluster.local:8080"
+          }
+          # OpenClaw proactive push
+          env {
+            name  = "TASK_WEBHOOK_URL"
+            value = "http://task-webhook.openclaw.svc.cluster.local"
+          }
+
+          readiness_probe {
+            http_get {
+              path = "/healthz"
+              port = 8080
+            }
+            initial_delay_seconds = 5
+            period_seconds        = 10
+          }
+          liveness_probe {
+            http_get {
+              path = "/healthz"
+              port = 8080
+            }
+            initial_delay_seconds = 30
+            period_seconds        = 30
+          }
+
+          resources {
+            requests = { cpu = "100m", memory = "384Mi" }
+            limits   = { memory = "768Mi" }
+          }
+        }
+      }
+    }
+  }
+
+  lifecycle {
+    ignore_changes = [spec[0].template[0].spec[0].dns_config] # KYVERNO_LIFECYCLE_V1
+  }
+
+  depends_on = [
+    kubernetes_manifest.external_secret,
+    kubernetes_manifest.db_external_secret,
+  ]
+}
+
+resource "kubernetes_service" "recruiter_responder" {
+  metadata {
+    name      = "recruiter-responder"
+    namespace = kubernetes_namespace.recruiter_responder.metadata[0].name
+    labels    = local.labels
+  }
+
+  spec {
+    type     = "ClusterIP"
+    selector = local.labels
+
+    port {
+      name        = "http"
+      port        = 8080
+      target_port = 8080
+    }
+  }
+}
diff --git a/stacks/recruiter-responder/terragrunt.hcl b/stacks/recruiter-responder/terragrunt.hcl
new file mode 100644
index 00000000..08c1ee48
--- /dev/null
+++ b/stacks/recruiter-responder/terragrunt.hcl
@@ -0,0 +1,24 @@
+include "root" {
+  path = find_in_parent_folders()
+}
+
+dependency "platform" {
+  config_path  = "../platform"
+  skip_outputs = true
+}
+
+dependency "vault" {
+  config_path  = "../vault"
+  skip_outputs = true
+}
+
+dependency "external-secrets" {
+  config_path  = "../external-secrets"
+  skip_outputs = true
+}
+
+inputs = {
+  # Override per-deploy in CI / commit. Initial build will land on forgejo
+  # as `forgejo.viktorbarzin.me/viktor/recruiter-responder:<8-char-sha>`.
+  image_tag = "latest"
+}